An Architecture for Mission-Bound Authorization

1.1. Map of This Document

The body reads in four parts, followed by appendices:¶

• The model comes first: a Mission's life end to end (Section 3); the Mission (Section 4), with its approval and lifecycle (Section 4.1), authority path (Section 4.2), delegated-authority layer (Section 4.3), capability envelope (Section 4.4), and survivable incorrectness (Section 4.5); what the family does not do (Section 5); the roles (Section 6); the verb spine that organizes every mechanism (Section 7); the invariants (Section 8); and who owns meaning and who commits authority (Section 9).¶

• The substrate and the bindings follow (Section 10): the binding-neutral kernel, the validity model (Section 10.4), and how the bindings differ (Section 10.5).¶

• Assurance (Section 11) covers what a deployment adopts (Section 11.1), what it can prove (Section 11.3), how its paths are bound (Section 11.4), and what a kill reaches (Section 11.5, Section 11.6).¶

• Deployment (Section 12) covers entry ramps (Section 12.1), patterns, and the Deployment Profile (Section 12.6).¶

• The appendices hold the requirements the family answers (Appendix A), a comparison with a conventional stack (Appendix B), a worked composition (Appendix C), the error surfaces (Appendix D), illustrative verification guidance (Appendix E), the Deployment Profile shapes (Appendix F), and the document map (Appendix G). DRAFTS.md in the repository is the full catalog.¶

Readers with one goal can start on a shorter path:¶

• New to Missions: Section 3, Section 4, and Section 7.¶

• Choosing what to deploy: Section 11 and Section 12.¶

• Writing or assessing a binding: Section 10, Section 11.4, and Appendix A.¶

• Reviewing security: Section 8, Section 11.5, and Section 13.¶

4.1. Approval and Lifecycle

In every binding, an explicit approval event creates the Mission and commits its approved context, or a verifiable commitment to it, and the Mission's Controller then owns its state. In the OAuth binding, a client proposes a Mission Intent, and may propose concrete authority alongside it; the Mission Issuer derives an Authority Set; and the approval event commits both.¶

In the OAuth binding, the commitment is the integrity anchors:¶

• intent_hash over the approved Mission Intent;¶

• authority_hash over the consented Authority Set; and¶

• where the client submitted an authority proposal, proposal_hash over the submitted authorization_details array.¶

Each anchor is computed over a domain-separated, issuer-bound envelope with fixed canonicalization, so an auditor can reproduce each digest from the record alone (the OAuth binding's Mission Approval, Integrity Anchors, and Canonicalization Rules sections). The record is immutable except for its state (the Mission Record section).¶

In the OAuth binding the lifecycle states are active, revoked, and expired, and only active permits issuance or a new positive governance decision. A non-active state stops further derivation and refresh at once. A credential already issued ends at the earliest of an applicable revocation, a state-aware or runtime check that reaches it, or its own expiry (Section 10.4).¶

Companions add states (suspended, completed, superseded, cascaded). One rule keeps these additions safe without a registry: a consumer treats every state other than the exact value active, including one it does not recognize, as non-active, so an unrecognized state fails safe (the OAuth binding's Mission Lifecycle and Gating section).¶

AAuth has its own counterparts to the OAuth commitment and lifecycle. Its exact-byte s256 commits the private approved mission blob. The approving PS and s256 are the Mission Reference, AAuth's mission identity (Section 8.2.1 of [I-D.draft-hardt-oauth-aauth-protocol]). On the wire the reference is the mission_s256 claim or parameter. The blob carries no member naming the PS: a person token names it by iss, a resource or auth token by ps, and a request by its destination.¶

In AAuth, the PS applies contextual governance using that blob and the ordered mission log; scopes, resource tokens, Resource and Access Server policy, and optionally AAuth Rich Resource Requests (R3, [I-D.draft-hardt-aauth-r3]) carry deterministic resource authorization. AAuth does not add the OAuth Authority Set or its two anchors, and its native lifecycle remains exactly active or terminated.¶

4.2. The Authority Path

In OAuth's names, authority moves along one path from the approval event to an action, and no stage holds more than the stage before it:¶

 approval basis
       |
       | adjudication
       v
 approval event
       |
       v
 Approved Authority Set (anchored, immutable)
       |
       | containment, discharge: subtract only
       v
 Effective Authority Set --child creation--> new approval event
       |               |                     for a Child Mission
       | derivation:   |
       | subset rule   | current set
       v               |
 derived credential    |
       |               |
       v               v
 runtime decision, where deployed
 (with action parameters and lifecycle state)
       |
       v
 action

Table 2 places each mechanism on this path.¶

The lifecycle gate sits beside the path, not on it: only active permits issuance, refresh, and a new positive governance decision, and resume reopens a suspended Mission's gate without changing either set ([I-D.draft-mcguinness-oauth-mission-status]). Widening and restoration never happen in place; a new approval event creates a successor ([I-D.draft-mcguinness-oauth-mission-expansion]).¶

The runtime decision takes the current Effective Authority Set as its own input, so narrowing applied after issuance binds there ([I-D.draft-mcguinness-mission-runtime]). Where no runtime decision is deployed, a resource relies on the credential alone, and the materialized-credential residual remains (Section 10.4).¶

4.3. The Delegated-Authority Layer

This family defines a delegated-authority layer. It governs the approved task a delegate performs on a principal's behalf: its bounded authority, its lifecycle, the per-action check, and the evidence that binds back to it. Authentication and token issuance answer who is acting and what a single credential carries; governance of standing entitlements answers what a principal should hold over time. Neither governs that task. This layer composes with the layers below rather than replacing them.¶

The Mission Authority Server ([I-D.draft-mcguinness-mission-authority-server]) is the layer's binding-independent control plane across an estate, whichever party issues a given token. Within the layer:¶

• the issuer side holds desired state (the record, its authority, its lifecycle and state version), reconciles it (gating, the ceiling review, evidence reconciliation), and distributes bounded authority (policy views, status, the management surface);¶

• tokens and the PEP/PDP boundary are the layer's data plane; and¶

• Status and Signals are the channel between the two.¶

The plane the layer governs is authority, never operations: how an agent runs stays with the harness and the orchestrator.¶

Grouped as planes, the roles of Section 6 form this layer, with the evidence surface crossing all of them:¶

 control       Mission control point (OAuth AS | MAS | AAuth PS |
               UMA AS | GNAP AS):
               approved context, lifecycle and gating;
               anchors and authority distribution where defined
                    |                       ^
                    | state, authority      | evidence
                    v                       |
 enforcement   PEP and PDP: a permit per consequential
               action, parameter binding, custody
                    |                       ^
                    | mediated actions      | outcomes
                    v                       |
 execution     harness, agent, orchestrator: sessions,
               sub-agents, queues, unwinding

 evidence      Consent Evidence, decision and execution
 (crossing)    evidence, Mission Receipts, the Mandate,
               audit transparency

4.4. The Capability Envelope

A Mission commits its authority and intent once, at approval, but an agent's work is open-ended: the actions a task will take are not known when it is approved. "Reconcile Q3 invoices" must authorize reading any invoice and posting any adjustment under the cap, because the specific ones cannot be enumerated in advance.¶

So the Authority Set an Approver consents to is a capability envelope, not a task specification, and the gap between that envelope and what a given run actually does is where agent risk lives.¶

The family's mechanisms are levers that narrow that gap:¶

• constraint-bounding and the subset rule shrink the envelope at issuance;¶

• runtime enforcement checks each action against it at the point of use;¶

• action-bound approval re-consents the highest-consequence actions with their concrete parameters;¶

• progressive authorization trades many increment approvals for one bounded ceiling;¶

• metering caps cumulative consumption; and¶

• completion retires authority as the task finishes.¶

No single lever closes the gap; a deployment composes the ones its risk warrants (Section 11), and the verbs of Section 7 organize the levers by the question each answers.¶

In the OAuth binding, what approval commits is broader than the structured Authority Set alone ([I-D.draft-mcguinness-oauth-mission]); it also commits:¶

• the rendered intent context (goal, task_bounds, and, where it differs, the requested ceiling);¶

• the effective expires_at; and¶

• the rendered derivation_limit and any metering bound ([I-D.draft-mcguinness-oauth-mission-derivation-limits], [I-D.draft-mcguinness-mission-metering]).¶

Concrete request values, current consumption, and action sequencing are decision-time facts, evaluated later by runtime policy, metering, or action-bound (transaction) approval. The OAuth binding does not require them to be re-rendered to the original Approver, though action-bound approval may re-render exactly that.¶

The lifecycle control gates new derivation from the envelope; a credential already materialized under it keeps running to its own bound unless an action-time gate reaches it first (Section 10.4, Section 11.6). Approval-time commitment and decision-time checks compose but do not substitute: an envelope real at approval time can still admit, at decision time, an effect the Approver never saw rendered in that form.¶

Attribution has the same limit at task-item grain: proving which concurrent task item produced a permitted action needs a verified cross-link that no family carrier supplies (Section 11.4).¶

4.5. Survivable Incorrectness

The levers share one strategy: they convert semantic risk into structural signals. A policy decision point is never asked to judge whether content is harmful. Provenance (the harness taint context), composition (the quarantine pattern), egress enumeration and volume bounds, separation of duty, and re-consent turn that question into facts a decision can gate on. A content evaluator a deployment adds composes as Resource policy at the decision point and only ever narrows.¶

The stance beneath the levers is survivable incorrectness: the agent is probabilistic, so the family never bets on the model being right and builds so that wrong is survivable, on two arms:¶

• The action arm is the capability envelope (Section 4.4) and its levers, wire-backed, with in-flight work unwinding through recorded reversibility classes ([I-D.draft-mcguinness-mission-orchestration]).¶

• The input arm is least exposure: everything the agent sees can steer it and everything it holds can leak, so a Mission budgets disclosure as well as authority, and mediated custody generalizes from credentials to context.¶

The arms differ in maturity. The input arm's enforceable edges are the harness taint rule, egress mediation, and catalog filtering ([I-D.draft-mcguinness-mission-harness], [I-D.draft-mcguinness-mission-runtime], [I-D.draft-mcguinness-mission-authzen]). Its interior (retrieval, memory, and context assembly scoped to the Mission) has no interoperable form and is deployment discipline, declared in the Deployment Profile (Section 12.6) rather than claimed.¶

6.1. The Actor Chain

One material action splits across the roles below. The family keeps each distinct and attributable rather than collapsing them into one "agent" identity. Each binding carries the same distinctions in its own vocabulary; AAuth does so natively with its agent identifier, the Person at the PS, parent_agent for a parent-mediated sub-agent, and the call chain for a service hop ([I-D.draft-mcguinness-mission-aauth]). The entries below show the OAuth binding's instantiation:¶

Principal:

the Subject, the token sub (the OAuth binding).¶

Accountable approver:

the Approver, committed at the approval event (the OAuth binding).¶

Intent generator:

the shaper, with Shaping Evidence recording what it emitted ([I-D.draft-mcguinness-mission-shaping]).¶

Authorizer:

the Mission Issuer at issuance (the OAuth binding); the PDP per action ([I-D.draft-mcguinness-mission-runtime]).¶

Approved agent:

the OAuth client, client_id on every derived token (the OAuth binding).¶

Executing delegate:

the outermost act actor (the OAuth binding's Delegation section).¶

Credential holder:

the mediating PEP under mediated custody ([I-D.draft-mcguinness-mission-runtime]).¶

Capability executor:

the executor of the capability source binding ([I-D.draft-mcguinness-mission-capability-binding]).¶

Downstream identity:

the audience-scoped token and the cross-domain local subject ([I-D.draft-mcguinness-oauth-mission-cross-domain]).¶

Attribution survives because no role is inferred from another: each is carried by its own construct, and the evidence layer records them together, in runtime evidence and the Mission Receipt ([I-D.draft-mcguinness-mission-runtime]), Consent Evidence, and the audit feed.¶

The chain is actor lineage, not authority lineage. An act chain records who acted through whom; it does not carry what task was approved, how authority narrowed at each derivation, whether the task remains active, or which parameter constraints bind. Those travel in the Mission's own constructs: the anchors, the Authority Set, the lifecycle state, and Child Mission lineage ([I-D.draft-mcguinness-oauth-mission-child-delegation]). Reading an actor chain as authorization provenance is the gap the Mission's lineage exists to close.¶

6.2. Agent Identity, Agent Deployment, and Mission

A deployment that runs agents under both an agent identity system and this family governs three independently lifecycle-bearing objects, separate from the authorization flow's four objects (Section 4). Each has its own owner, lifecycle, and revocation, and the model stays clean only while none absorbs another's job:¶

Agent identity (who is acting):

The logical agent and, where client instance identification is deployed, the concrete instance. Owned by the deployment's agent IAM, a registry or directory outside this family, and consumed, in the OAuth binding, as the client_id, the client-instance attestation, and validated Instance Context.¶

A deployment may authenticate concrete agent instances with client instance identification, which extends attestation-based client authentication ([I-D.draft-ietf-oauth-attestation-based-client-auth]), and endorse its attesters in client metadata ([I-D.draft-mcguinness-oauth-client-instance-id], [I-D.draft-mcguinness-oauth-client-attesters]). Doing so sharpens joins and evidence attribution to instance granularity without touching the Mission model. Instance evidence alone identifies no actor or delegate ([I-D.draft-mcguinness-oauth-client-instance-id], Section 5). SPIFFE workload credentials ([I-D.draft-ietf-oauth-spiffe-client-auth]) on their own convey no instance identity under that profile ([I-D.draft-mcguinness-oauth-client-instance-id], Section 8.2). None of these mechanisms defines the Mission actor, delegation, intent, lifecycle, or evidence semantics defined here.¶

Agent Deployment (what is running):

The approved behavioral version of the agent: its code, model, system prompt, tool allowlist, data scope, and runtime configuration. Owned by the deployment's change governance. A change to any of these is a new Agent Deployment; which changes require re-approving standing Missions is policy that governance records.¶

A Mission may be pinned to a named Agent Deployment where a companion Agent Deployment Binding profile realizes that pin; [I-D.draft-mcguinness-oauth-mission] names the profile and its required properties but defines none of its wire mechanics itself. An Agent Deployment is distinct from the Mission Deployment Profile (Section 12.6), the estate's published claims manifest, not a property of an agent.¶

Mission (why the authority exists):

This family's object: the approved task, its lifecycle, and, where the binding derives one, its Authority Set.¶

An agent registry is a complementary dependency, not part of the Mission system; an A2A AgentCard directory ([A2A]) is one example source. Where one exists, the Mission Issuer and the PDP consume a small, stable slice of it:¶

• the agent identifier and its owner,¶

• current status and revocation state,¶

• the approved Agent Deployment,¶

• eligibility bounds (what the registry permits the agent to be approved for, a derivation input, never a grant), and¶

• risk tier.¶

Discovery contents are not authority for workload identity, authorization, or effective capabilities: a deployment applies local admission policy and independently verifies the relevant credential, resource metadata, and capability evidence.¶

Registry state is a state source like any other: the consuming decision point treats it under the runtime profile's freshness discipline, with a declared staleness bound, failing closed when it cannot be established ([I-D.draft-mcguinness-mission-runtime]).¶

Authorization composes conjunctively across these lifecycles and the credential's own: a decision may depend on agent state, Mission state, and credential validity, and each gates independently. A valid credential never overrides a revoked agent or a non-active Mission, and a live agent under an active Mission still fails on an expired credential.¶

Binding strength accumulates as a deployment adds it:¶

• authority is issued to an authenticated client;¶

• sender-constraint keys pin possession;¶

• client-instance attestation over an instance-unique key pins the concrete instance;¶

• attested runtimes pin the execution environment; and¶

• an Agent Deployment pin holds the behavioral version.¶

The key and instance steps are the presenter-key-bound and instance-bound binding properties (Section 11.4); none of these steps is an assurance level.¶

In the division of labor with agent IAM, agent identity preserves who is acting, and the Mission preserves why their authority exists:¶

• the registry and workload identity authenticate an approved agent instance;¶

• the Mission and, where the binding derives one, its Authority Set say what sanctioned work that instance carries;¶

• per-hop credentials narrow;¶

• the runtime layer enforces each action and parameter; and¶

• the evidence layer joins what was approved, decided, and done.¶

7.1. Propose

Question:

How does a user's request become a candidate approved task?¶

Boundary:

The client side, in OAuth.¶

Owners:

• Intent Shaping, which produces an untrusted Mission Intent ([I-D.draft-mcguinness-mission-shaping]); and¶

• the submission-evidence framework, under which any Intent Submission Evidence is processed ([I-D.draft-mcguinness-oauth-mission-submission-evidence]).¶

In OAuth, the Intent enters through Pushed Authorization Requests [RFC9126] or the MAS submission endpoint. In AAuth, the agent sends the native description and requested tools to the Person Server's mission endpoint. The AAuth binding defines no Mission Intent or dependency on the shaping profile.¶

7.2. Approve and Record

Question:

How does a proposed task become an approved, committed Mission?¶

Boundary:

The binding's control point; the approval event is where trust is created.¶

Owners:

• the five bindings ([I-D.draft-mcguinness-oauth-mission], [I-D.draft-mcguinness-mission-authority-server], [I-D.draft-mcguinness-mission-aauth], [I-D.draft-mcguinness-mission-uma], [I-D.draft-mcguinness-mission-gnap]);¶

• Consent Evidence, committing the disclosure shown to the Approver ([I-D.draft-mcguinness-oauth-mission-consent-evidence]); and¶

• Deferred Approval ([I-D.draft-mcguinness-oauth-mission-approval]), the OAuth binding's asynchronous path, with an experimental companion adding an in-review narrowing negotiation.¶

The standalone and AAuth bindings are natively asynchronous. AAuth's baseline approval commits its exact mission blob, not the family Consent Evidence object. Where the experimental progressive authorization companion is used, the initial approval also consents an authority ceiling for later staged widening ([I-D.draft-mcguinness-oauth-mission-progressive]).¶

Who holds the deciding side is a spectrum, not a species. The approval event requires an accountable principal deciding against committed inputs before any authority exists, and the proposer is never the approver.¶

A deterministic, versioned policy can approve at machine speed within a ceiling a human consented to (the template, policy_drawdown, and ceiling_drawdown authorization bases that companion profiles define on the OAuth binding's extension point; see Authorization bases in Section 8). Policy approves the instance because a human approved the policy or the template; policy_version keeps that chain re-checkable. Derivation fixes the authority, the adjudicator (a human or such a policy) decides activation, and a human is accountable.¶

A model's judgment, over risk signals or enterprise context, enters adjudication only as a recorded input to the policy: it can refuse or narrow, never grant or widen. Neither the policy nor a model input to it gates on the Intent's prose members, which stay the human Approver's check ([I-D.draft-mcguinness-oauth-mission]). A generated approver reading attacker-influenced proposals is itself an injection surface. The high-consequence classes stay on a fresh human decision, per the progressive profile's prohibited set ([I-D.draft-mcguinness-oauth-mission-progressive]).¶

7.3. Govern

Question:

How do consumers observe Mission state, and how does authority grow or retire mid-task?¶

Boundary:

Between the issuer and every consumer relying on state.¶

Owners:

• Status, the signed pull surface with a lifecycle endpoint ([I-D.draft-mcguinness-oauth-mission-status]), extended by the Status List companion for fleet-scale reliance ([I-D.draft-mcguinness-oauth-mission-status-list]) and by the Entry Discharge companion for per-entry completion discharge ([I-D.draft-mcguinness-oauth-mission-discharge]);¶

• Signals, the push complement ([I-D.draft-mcguinness-oauth-mission-signals]);¶

• Expansion, widening only via an approved successor ([I-D.draft-mcguinness-oauth-mission-expansion]);¶

• Containment, event-triggered monotonic narrowing of a live Mission's effective authority ([I-D.draft-mcguinness-oauth-mission-containment]);¶

• Management, fleet enumeration and bulk lifecycle for operators ([I-D.draft-mcguinness-oauth-mission-management]); and¶

• Discovery, experimental, binding encountered resources within a pre-consented ceiling ([I-D.draft-mcguinness-mission-discovery]).¶

The AAuth binding carries this verb natively through its management companion for status, termination, and delegation-tree queries ([I-D.draft-mcguinness-mission-aauth-management]). Its native approved lifetime bound is AAuth's own expires_at, profiled by the expiry document ([I-D.draft-mcguinness-aauth-mission-expiry]).¶

7.4. Enforce Each Action

Question:

Is this specific action, with these parameters, permitted under this Mission now?¶

Boundary:

The last controllable point between agent and resource.¶

Owners:

• the runtime profile, the decision contract with parameter binding, custody, and fail-closed behavior ([I-D.draft-mcguinness-mission-runtime]);¶

• its AuthZEN profile, the concrete decision API ([I-D.draft-mcguinness-mission-authzen]); and¶

• the runtime evidence companion, the Decision Evidence, Execution Evidence, and Refusal Record objects ([I-D.draft-mcguinness-mission-runtime-evidence]).¶

The runtime decision composes conjunctively with the other gates. Each of the following is independently necessary, and none grants, widens, or restores another ([I-D.draft-mcguinness-mission-runtime], Section "The Runtime Decision"):¶

• Effective Authority Set membership;¶

• every applicable cumulative-consumption or stateful operational gate; and¶

• a required action-bound approval.¶

7.5. Run and Wind Down

Question:

How does governed work start, persist, pause, and unwind when Mission state changes?¶

Boundary:

The operator's execution environment around the agent.¶

Owners:

• the harness, binding session continuity to Mission state ([I-D.draft-mcguinness-mission-harness]); and¶

• Orchestration, unwinding in-flight work through reversibility classes and recorded unwind plans ([I-D.draft-mcguinness-mission-orchestration]).¶

7.6. Delegate

Question:

How does authority reach a sub-agent without widening?¶

Boundary:

Between principals acting under one approval.¶

Owners:

• Child Delegation, child Missions with lineage, strict-subset authority, and cascade revocation ([I-D.draft-mcguinness-oauth-mission-child-delegation]); and¶

• Offline Attenuation, narrower Mission-bound tokens minted off the issuer's hot path ([I-D.draft-mcguinness-oauth-mission-attenuation]).¶

Both build on the actor chain of the OAuth binding's Delegation Within a Mission section. Offline attenuation requires the runtime enforcement layer: its kill switch is the runtime state re-check.¶

The chooser:¶

• the OAuth binding's token-exchange delegation ([RFC8693]), for an execution hop that lives and dies with the parent's lifecycle;¶

• a Child Mission, when the delegate needs its own lifecycle, approval, or audit identity; and¶

• offline attenuation, experimental, only where offline minting is the constraint.¶

The same principal exercising the same authority concurrently is not delegation but swarm execution (Section 7.7).¶

AAuth delegates natively: a parent-mediated sub-agent under parent_agent, distinct from the call chain of a service hop, with no Authority Set machinery imported ([I-D.draft-mcguinness-mission-aauth]).¶

7.7. Swarm Execution: Multiplication, Not Delegation

Swarm execution composes the three lifecycle-bearing objects (Section 6.2): one Mission, executed concurrently by N instances acting under the same authorized agent identity. Where the deployment implements the Mission's Agent Deployment pin, every instance also satisfies it. It is multiplication, not delegation: no act hop, no Child Mission, no attenuation chain, because authority never moves between principals.¶

The Agent Deployment pin is a named architectural pattern, not a wire member the OAuth binding defines. That document reserves no Intent member for it and points to a dedicated Agent Deployment Binding profile, which a deployment wanting the pin implements ([I-D.draft-mcguinness-oauth-mission]). An agent identity and an Agent Deployment are distinct objects (Section 6.2): a profile may map a Deployment version to a distinct client registration, but this document neither requires nor implies that mapping.¶

The invariant is shared-identity authorization with per-instance attribution. Authorization attaches to the shared agent identity, and that shared identity does not distinguish individual instances. For deployments using OAuth client authentication and the client-instance profile, including applicable MAS joined paths, late binding is attestation: an instance joins the work by authenticating as the shared client with its own Client Attestation ([I-D.draft-mcguinness-oauth-client-instance-id]), not by receiving a credential from a peer. Attribution then stays per-instance through verified Instance Context, which attributes a presentation to an instance only under a sender-constraint key unique to it ([I-D.draft-mcguinness-oauth-client-instance-id], Section 7.3). These mechanisms realize per-instance attribution; they do not themselves define the Deployment-version pin.¶

The derivation limits profile's derivation_limit ([I-D.draft-mcguinness-oauth-mission-derivation-limits]) is not a fan-out or concurrency ceiling. It counts issuance events, not instances: a single instance refreshing N times consumes it exactly as N instances deriving once each would. What bounds a swarm's aggregate consumption is the metering profile's Mission-grain budget: consumption bounds attach to the Mission, not to any one instance, so a swarm of instances shares one budget ([I-D.draft-mcguinness-mission-metering]).¶

Swarm execution comes before the Delegate verb's chooser: more instances under the same authorized agent identity deriving under one Mission, with no new construct. A different principal acting inline, a durable sub-agent, and offline narrowing take the chooser's options.¶

7.8. Project

Question:

How is one Mission honored in another trust domain?¶

Boundary:

A trust boundary the origin does not control, where the verifier holds no session with the issuer.¶

Owner:

Cross-Domain Projection, a single-hop grant that carries the Mission's identifier, issuer, and authority hash into a partner domain unchanged, where a Resource AS mints a local token bounded by the projected authority ([I-D.draft-mcguinness-oauth-mission-cross-domain]).¶

Projection preserves authority across the boundary rather than narrowing it to a sub-actor, which makes it a distinct verb from Delegate. Downstream revocation latency is the local token lifetime.¶

7.9. Continue

Question:

How does a Mission's authorization continue, under the same approval and constraints, when the acting identity must be re-established at each hop or after the original credential is gone?¶

Boundary:

The seam between authorization continuity and identity continuity.¶

Owner:

Mission Continuation ([I-D.draft-mcguinness-oauth-mission-continuation]), the authorization-continuity profile.¶

The profile keeps three easily conflated things apart:¶

• Identity continuity (who is acting, and how that identity legitimately continues) rides a transport, not this profile: Identity Continuation ([I-D.draft-mcguinness-oauth-id-continuation-assertion]) for short-lived, sender-constrained hops among Resource Authorization Servers that trust a common identity provider; async delegation ([I-D.draft-zhu-oauth-async-delegation]) for a long-running, disconnected task; and the cross-domain grant ([I-D.draft-mcguinness-oauth-mission-cross-domain]) across a trust boundary.¶

• Authorization continuity (what work remains authorized, under which constraints, on whose approval) is the Mission's. Every continued grant derives a subset of the Mission's Authority Set, is state-gated at issuance, and is bounded by the Mission's expiry. When the Mission goes terminal, new continued grants end; authority already issued under a continued grant runs to its own bound unless a state-aware gate reaches it first.¶

• Execution-time evidence records, against the Mission, what was done at each continued hop.¶

A continuation handle grants nothing: it names an accepted hop, and every continued grant re-passes the Mission's active gate. Continuity is never authority, the rule the harness applies to session continuity ([I-D.draft-mcguinness-mission-harness]).¶

Continue differs from Delegate, which narrows authority to a sub-actor. Continue uses the Project verb's cross-domain grant as one transport rather than replacing it.¶

7.10. Prove

Question:

What can a party outside the deployment verify about what was approved and done?¶

Boundary:

Across trust domains and time; the verifier holds no session with the issuer.¶

Owners:

• Consent Evidence ([I-D.draft-mcguinness-oauth-mission-consent-evidence]);¶

• the Mandate, a signed, portable statement that authorizes nothing ([I-D.draft-mcguinness-mission-mandate]);¶

• the Mission Receipt, portable evidence of an action taken under a Mission ([I-D.draft-mcguinness-mission-runtime]); and¶

• Audit Transparency, the append-only evidence log ([I-D.draft-mcguinness-mission-audit]).¶

7.11. Analyze

Question:

Which components must be trusted, and what does each one's compromise cost?¶

Boundary:

The whole system.¶

Owner:

The Mission Security Model ([I-D.draft-mcguinness-mission-security-model]).¶

9.1. The Ontology Contract

The capability envelope (Section 4.4) meets its hardest case in the open world. The OAuth model starts with authority the client proposes and the approval enumerates. An agent that discovers resources at encounter time breaks that premise, and some authorization mechanisms invert it: the resource declares its own operations and consequences.¶

The derivation boundary (Section 9.2) settles who commits authority; this section settles who owns what an operation means. The resource owns the ontology, its operations, its constraint semantics, and their consequences. Derivation, consent rendering, and enforcement consume that meaning without owning it, and no layer invents meaning it does not own. One boundary is shared by agreement: the resource owns its operation semantics and consequences, while the family owns the registered cross-resource constraint vocabulary, which a resource explicitly advertises and adopts before it binds ([I-D.draft-mcguinness-oauth-mission-resource-access]).¶

The consuming contract: meaning binds at approval and is enforced at the point of use. Any translation between the resource's vocabulary and another party's is trusted, verified, or separately approved, never a place where authority widens.¶

Resource-owned meaning reaches the three consuming layers through five mechanisms, each defined in its own home and composing as one contract:¶

Common Constraints:

The registered constraint vocabulary every conforming party evaluates identically, with the mission_constraints_supported protected-resource metadata member advertising which constraints a resource enforces. Home: the OAuth binding's Mission Resource Access Profile ([I-D.draft-mcguinness-oauth-mission-resource-access]).¶

Capability-source binding:

Catalog-sourced capability definitions (an MCP tool, an OpenAPI operation) content-digested at derivation and refused on drift at decision time, so the meaning authority bound to is the meaning enforced. Home: the capability-binding companion ([I-D.draft-mcguinness-mission-capability-binding]).¶

Operation Profiles:

The per-operation statement of normalization and binding rules, carrying resource-declared operation semantics such as idempotency, reversibility, and lease requirements into parameter binding. Home: the runtime profile ([I-D.draft-mcguinness-mission-runtime]).¶

The encounter contract:

What is submitted, adjudicated, and recorded when an agent meets a resource the approval could not enumerate, so meaning that arrives late still binds before use. It covers the encounter's routing (through drawdown, catalog binding, projection, or fresh approval), identity pinning, and floors. Home: the discovery companion ([I-D.draft-mcguinness-mission-discovery]).¶

Resource-Declared Semantics:

The full inversion: the resource publishes its operations, their human meaning, and their consequences. Under the OAuth discovery composition, the declaration can be content-addressed by r3_s256 as an additional commitment beside the Mission's integrity anchors, recording what the resource claimed to be when authority bound to it. In that composition, the declared operations become candidate vocabulary that derivation narrows against. AAuth can instead use R3 as a resource-owned deterministic authorization vocabulary while the private mission blob remains contextual PS governance. R3 content addressing is not a baseline AAuth mission anchor. Home: the discovery and R3 compositions, informative.¶

Behind the five mechanisms sits one direction axis, chosen per encounter, not fixed by binding. The family inherits OAuth's client-proposed default: the client names the authority it wants, and the resource's meaning arrives through metadata, catalogs, and profiles. Resource-Declared Semantics is the inversion, where the resource speaks first. Where a structured-authority binding commits that meaning at approval, it is enforced at use and translation never widens.¶

Under the OAuth binding, the resource-declared direction runs entirely through existing family seams:¶

• the encounter contract routes the declaration ([I-D.draft-mcguinness-mission-discovery]);¶

• narrowing-mode derivation consumes the declared operations as candidate vocabulary ([I-D.draft-mcguinness-oauth-mission]);¶

• consent composes the resource-authored material; and¶

• the declaration's digest rides the derived authority (the progressive companion's resource_declaration_digest, [I-D.draft-mcguinness-oauth-mission-progressive]).¶

The OAuth-native descriptive surfaces the resource-declared direction builds on are an authorization server's RAR type metadata, the schema and documentation of each type it supports ([I-D.draft-ietf-oauth-rar-metadata-remediation]), and a protected resource's authorization_details_types_supported ([RFC9728]).¶

In AAuth, R3 can describe deterministic resource authorization independently of mission approval; the PS considers the resource request and mission context without turning the R3 declaration into a Mission Authority Set.¶

Where a deployed semantic-binding mechanism is in force, both directions close the same loop: the meaning source's digest becomes part of the derived authority. A catalog-sourced capability pins its source_digest, and a resource declaration pins r3_s256. In each case the Authority Set carries the version of the meaning it was derived under, and the point of use compares against the meaning in force. Under such a mechanism, meaning is not consulted at approval and assumed at enforcement; it is committed at approval and re-verified at use. An ordinary registered authorization_details type can carry stable semantics with neither digest; the loop is closed by the mechanism a deployment runs, not by the family universally.¶

Meaning, like state, fails closed. Each home states the contract's failure mode: a consumer facing an operation meaning it cannot resolve, a constraint it cannot evaluate, a drifted capability definition, or an unrecognized declaration refuses rather than guesses.¶

The contract has a dual of equal force. The resource owns what an action means; the Mission owns why it is happening and where the undertaking stands, and that context exists nowhere else in the stack. A resource evaluates each request at perfect local resolution and zero task resolution: it can price every consequence its ontology names but cannot see the undertaking the request belongs to. A risk decision composes both sides of this context asymmetry. Semantics without purpose prices every delete the same; purpose without semantics cannot read the call.¶

"Delete database" in isolation is indistinguishable from catastrophe. "Delete database" inside an approved migration whose copy steps already completed is a priced, checkable step. That judgment needs the undertaking's history, and no resource-local view contains it. After the fact, that history is reconstructible from the join of Decision and Execution Evidence on the Mission's identity. At decision time, a task-aware decision point can draw the same history from trusted prior workflow state or another authoritative source. The runtime profile names the mechanism: sequence-aware evaluation over the undertaking's history is an optional decision input, guarded so that history informs a decision and never widens one ([I-D.draft-mcguinness-mission-runtime]).¶

9.2. The Authority Derivation Boundary

In the authority-bearing bindings, the Approver consents to the Authority Set the Mission Issuer renders and commits, not to the Mission Intent. All of these bindings derive under the OAuth binding's rules ([I-D.draft-mcguinness-oauth-mission], Section "Mission Authority").¶

Derivation is mechanical: the Mission Issuer narrows a submitted authority proposal to policy, or looks up candidate entries in a configured mapping keyed on structured Intent members and narrows those. The prose members (goal, task_bounds, success_criteria) never change what is derived; they bound authority through disclosure, since the Approver refuses authority the words do not support. Translating a user's words into structure is the shaper's job, before admission and outside the trust boundary ([I-D.draft-mcguinness-mission-shaping]).¶

Derivation policy stays local: what the Mission Issuer narrows to, and what its mappings contain. The consequence is a trust boundary: interoperability begins at the committed result, not at the Intent. A Mission Intent has no portable semantics; two conforming Mission Issuers can derive different Authority Sets from the same submission. Audit can establish what was derived (against intent_hash and policy_version), never whether it was the right reading of the task.¶

Narrowing mode is the interoperable path: each derived entry is a subset of a proposed entry, so the result is reproducible and the client can check it against its proposal. A deployment whose partners must reason about its derivations can publish a derivation policy identifier and test fixtures that pin Intent-to-Authority-Set outcomes, making the local policy reviewable even though it does not travel. The OAuth binding's informative Derivation Policy appendix ([I-D.draft-mcguinness-oauth-mission]) illustrates that policy artifact: a worked narrowing rule, fixtures, and ownership.¶

The ceiling the derivation narrows against is a composition, not a single object. The derived Authority Set sits inside every bound on the task:¶

• the issuer's derivation policy;¶

• the ceiling of the Mission's established authority source; and¶

• at enforcement time, the resource owner's and deployment's live policy at the decision point.¶

The derivation step intersects the first two and commits the result. The runtime contract re-checks the rest on every action, which is why a permit is never implied by the Authority Set alone. A deployment adding further sources (a tenant boundary, an environment-specific floor) adds them as derivation-policy inputs or as decision-point policy, never as agent-negotiated widening.¶

The established authority source is a delegating person's own authority, a workload's provisioned authority, or governed organizational policy. Approval activates authority the source already holds and grants nothing beyond it. The Approver needs authority to activate the source, not personal possession of its permissions.¶

The derivation modes rank by how portable their result is:¶

A deployment seeking interoperable authority uses narrowing; a model-assisted extension stays local policy unless a profile pins it with a published policy identifier, version, and test fixtures.¶

9.3. Approval Fidelity

For the portable-authority bindings, the approval event authenticates the Approver, establishes the Subject, derives and renders the Authority Set for consent, and computes the anchors over the consented set, the approved Intent, and, where one was submitted, the authority proposal. That approval event creates the record in active atomically with the decision.¶

AAuth approval has different fidelity: the native propose, clarify, and approve interaction authenticates the parties, returns the approved mission blob and exact-byte s256 commitment, and creates an active Mission Context. It does not render or commit an OAuth Authority Set.¶

Home: the OAuth binding's Mission Approval section. Consumed by Consent Evidence, which binds to this event, and by every downstream guarantee that assumes the anchors, the gating, and the record.¶

10.1. The Primitives, at a Glance

The anchors in the envelope row are commitment anchors, not enforcement proofs (Section 9.2). A Resource Server holding a narrowed token enforces the authority it receives rather than reconstructing authority from a hash of a full set it does not hold ([I-D.draft-mcguinness-mission-security-model]). A Resource Server or policy decision point that needs an independent check retrieves the complete approved set, recomputes its anchor, and checks the carried authority as a subset of it under Mission Approved-Set Verification ([I-D.draft-mcguinness-oauth-mission-approved-set-verification]).¶

10.2. Token Classes

"Mission-bound" is a specific claim. The OAuth binding's Terminology names three token shapes so that a weak one is not read as the strong one:¶

• a Mission-referenced token carries a Mission identifier only;¶

• a Mission-derived token carries authority derived from an active Mission; and¶

• a Mission-bound token is Mission-derived and additionally active-state gated, subset-constrained, and refresh-gated: in the contract's vocabulary, a credential covered by the binding's Lifecycle-Gated Authorization, Monotonic Derivation, and Credential-Bound claims.¶

Only the third earns the term: a mission claim alone is a reference, not Mission-bound authorization. The family defines "Mission-bound" by the credential-mission-bound equivalence of the binding properties (Section 11.4): six conditions, each mapped to what discharges it, earned only where all hold together. The OAuth binding discharges the equivalence through its own conformance rule.¶

10.3. The Mission-Bound Credential

The Mission-Bound Credential is a credential carrying the mission claim (id, issuer) and Mission-derived authorization details, issued only while the Mission is active. Its home is the OAuth binding's Mission-Bound Access Tokens and The Mission Claim sections. The claim's id and issuer are the Mission-identity condition; a token-carried Authority Set commitment is never required for it, per condition 2 of credential-mission-bound (Section 11.4).¶

The credential is the binding-dependent primitive, and the bindings split on it:¶

• the OAuth binding provides it;¶

• the standalone binding does not: per the MAS's Mission Substrate section, a MAS provides neither this credential nor issuance gating ([I-D.draft-mcguinness-mission-authority-server]); and¶

• an AAuth auth token can carry the native Mission Reference (mission_s256) but does not carry Mission-derived authorization details, so it is Mission-referenced, not a Mission-bound credential ([I-D.draft-mcguinness-mission-aauth]).¶

For profiles that compose with the credential, the seam is Mission binding establishment ([I-D.draft-mcguinness-mission-runtime], Section "Mission Binding Establishment"). It has two modes. Credential-carried: the acting credential carries the Mission reference, where the binding provides one. Externally established: the PEP supplies a reference that the PDP verifies against the acting credential under a join the binding defines, which the MAS profiles as its Mission Join; an unverified reference establishes no Mission.¶

Offline Attenuation attenuates this credential, and the token-carriage aspects of delegation ride it, so both require it. The companions that need a credential-to-Mission association (the runtime layer and the harness) route through Mission binding establishment, which is what makes the standalone binding possible.¶

The issuance-grant companion ([I-D.draft-mcguinness-oauth-mission-issuance-grant]) provides this primitive by composing the OAuth and standalone bindings: the standalone Mission Issuer mints a Mission Issuance Grant that a consuming Authorization Server redeems for Mission-bound tokens.¶

10.4. The Validity Model

Five validity horizons govern reliance, each with its own setter, checker, and consequence; implementations most often err by conflating them:¶

Token exp:

set by the credential issuer and checked by every consumer of the token. In the OAuth binding it is capped by the Mission's expires_at. Past it the credential is dead and obtaining a new credential re-enters whatever issuance gate the binding and access mode provide.¶

Mission state, and expires_at where defined:

set by the controlling authority and checked at the binding's declared control points. In OAuth, the issuance gate, PDP, and state consumers enforce it. In AAuth, the PS enforces active or terminated at PS endpoints and on the PS-gated modes; agent identity and resource-managed access have no PS state gate.¶

fresh_until:

set by the status responder; checked by status consumers. Past it a cached state report may not be relied on and is re-fetched.¶

Permit window:

set by the PDP; checked by the executing PEP. Past it the permit is void and a new decision is required.¶

Action-approval freshness:

set by the approval surface; checked by the PDP. Past it an action-bound approval no longer authorizes the action it named.¶

The horizons compose by minimum: reliance at any moment requires every applicable horizon to be open, and no horizon substitutes for another.¶

The horizons also give the deployment its freshness dial, whose TTL-only end is a first-class posture, not a fallback. A deployment that relies on lifetimes alone verifies with local cryptography and a clock: no state source, no freshness discipline, no availability coupling, and a worst-case exposure equal to the lifetime by construction. Where every fresh credential crosses a Mission-state decision point, this posture realizes the lifecycle-gated capability with reliance bounded by credential lifetime alone ([I-D.draft-mcguinness-mission-substrate]). That holds only for access modes with such a gate: AAuth's PS-gated modes have it, while agent identity and resource-managed access do not.¶

TTL-only is appropriate where the artifact's lifetime meets the tolerated exposure bound and no required state check would shorten that exposure. That is the case for short missions, and at action grain: an artifact there lives seconds to minutes, and a revocation landing inside its window has no observation point that could reach the artifact before its own expiry does. The family's short-lived artifacts (the permit, the cross-domain grant, the Join Assertion) sit at this end.¶

A lifetime cannot suspend, complete, or kill at once. The Mission's state carries that task-grain residue, and the state-observable capability (an authenticated freshness source with a stated staleness bound) reaches it. State-observable is a named substrate capability a binding may provide beyond the kernel's bounded-reliance floor ([I-D.draft-mcguinness-mission-substrate]), and it is the one runtime enforcement requires ([I-D.draft-mcguinness-mission-runtime]).¶

The two ends are one mechanism seen from opposite sides. A lifetime relocates the freshness check from the verification path to the issuance path, so every re-issuance is the policy re-check: the family's gates-new-derivation-only rule in its other reading. A deployment states where it sits on the dial through its bounded-revocation claim (Section 11.3): a TTL-only posture claims the lifetime as its bound only for paths whose re-issuance is gated, with no state-observable overlay. The runtime profile prices each position, source by source, in its state and freshness section ([I-D.draft-mcguinness-mission-runtime]).¶

10.5. Binding Security Architectures

The authority-bearing bindings are OAuth 2.0, the standalone Mission Authority Server, and (as experimental sketches) UMA 2.0 and GNAP. AAuth composes at the shared Mission Context layer: approval, stable reference, lifecycle gating where the PS is on path, and governance history (Section 4.1).¶

Peer standing among the bindings is a deployment-topology claim, not a data-model-independence claim: each attaches to its own protocol as an equally adoptable unit and declares what it supplies (Section 10). OAuth is the family's first-authored binding, built on widely deployed OAuth infrastructure, a deployment fact and not a maturity ranking. The Mission Authority Server ([I-D.draft-mcguinness-mission-authority-server]) is a normative standalone-controller protocol binding over the OAuth Mission data model: a peer deployment topology, but not an independent substrate model, since it normatively imports the OAuth Mission record and issuance profile. AAuth demonstrates model independence: it maps the shared kernel onto its own protocol's native Mission Context without importing the OAuth Authority Set ([I-D.draft-mcguinness-mission-aauth]). Peer standing implies neither identical capabilities nor identical adoption cost: adopting Missions on OAuth requires the changes the OAuth binding defines, while AAuth, natively contextual, adds no new wire members.¶

The bindings share Mission Context capabilities but are not one security system. Each has its own authority representation, trust assumptions, cutoff behavior, and failure modes, and a deployment names its architecture, not only its binding. Three patterns cover the bindings:¶

• credential-carried authority: the credential names the Mission, carries derived authority, and issuance is gated (the OAuth AS, UMA AS, and GNAP AS);¶

• PDP-joined: credentials are ordinary and a join establishes the association at the decision point (the standalone MAS); and¶

• context-carried: AAuth carries its native Mission Reference (mission_s256) while authority remains in resource scopes and policy. The PS-gated modes (Section 6) pass through the PS, while agent identity and resource-managed access bypass it.¶

The differences that decide a design:¶

A MAS deployment that holds the same Mission as an AS deployment does not thereby provide AS-native semantics, so a comparison of deployments compares their architectures first.¶

One Mission can govern authority issued by several authorization servers. Its Controller need not issue each credential, and each server may know the same participant by its own client identifier. The owning Mission Issuer records how those identifiers relate: the MAS in its mapping contract, and each issuance grant's client_id as the consuming AS knows it ([I-D.draft-mcguinness-mission-authority-server], [I-D.draft-mcguinness-oauth-mission-issuance-grant]). Identifier equality is therefore neither required nor sufficient for participant continuity. The kernel requires no client_id and no subset relation across administrative or protocol boundaries ([I-D.draft-mcguinness-mission-substrate]), and a client_id shared by several workloads is a join residual the MAS states.¶

The OAuth binding stacks two independent chokepoints. Issuance gating acts at the token layer: a revoked or expired Mission stops all further derivation and refresh, and short-lived tokens age out. Runtime enforcement acts at the action layer: each consequential action is re-checked against current state at the point of use. Issuance gating plus runtime enforcement is strictly stronger than either alone: a gap in PEP coverage is still bounded at the token layer, and an outstanding token is still stopped at the action layer on mediated paths.¶

The AAuth binding has a narrower structural chokepoint. The Person Server refuses new person-token issuance, PS authorization, or federated brokering for a terminated Mission Context, bounding those paths by token lifetime. Agent identity and resource-managed access do not cross that chokepoint. AAuth supplies the Mission Context capabilities in its own idiom but not a portable Authority Set or universal subset rule. The AAuth binding defines no generic family runtime composition or independently resource-verifiable Authority Set ([I-D.draft-mcguinness-mission-aauth]).¶

Per-action enforcement is budgeted, not blanket:¶

• only consequential actions are gated;¶

• the common-case decision is a local evaluation against a materialized policy view whose network cost is paid per freshness window; and¶

• the runtime profile requires a synchronous gate only for the high-consequence classes (the runtime profile's deployment considerations, [I-D.draft-mcguinness-mission-runtime]).¶

Runtime enforcement composes as an overlay, not a substrate swap: a deployment mediates the paths where the high-consequence classes live and lets every other resource ride lifetime-bounded reliance (Section 11.1), with token lifetimes sized to the tolerated staleness and no state evaluation at the resource.¶

The standalone mode trades the token-layer kill switch for zero Authorization Server changes. A MAS creates, approves, and serves Missions while tokens remain ordinary; the PDP joins credentials to Missions, and the MAS is the freshness source. The cost is structural:¶

• no mission claim travels;¶

• revoking a Mission stops nothing at the token layer; and¶

• enforcement rests entirely on PEP coverage, so a token exercised outside that coverage is ungoverned (the MAS's Limitations section).¶

The path to a token-layer chokepoint is the OAuth binding, where the estate's AS can change; the record, anchors, and lifecycle carry over unchanged. That is a move between peer architectures, not an upgrade from a lesser one: the MAS remains a peer binding, not a staging area (its own document's framing).¶

Between those two architectures sits the issuance join ([I-D.draft-mcguinness-oauth-mission-issuance-grant]): the MAS remains the Mission Issuer while estate Authorization Servers redeem Mission Issuance Grants for Mission-bound, state-gated tokens, restoring the token-layer chokepoint without moving approval.¶

11.1. The Four Levels

Each level includes the one before it:¶

Baseline Issuance:

the approved, anchored Mission record and its lifecycle: authority derived and committed at the approval event with the integrity anchors (the OAuth binding).¶

Credential-issuing bindings: where a structured-authority binding issues Mission-bound credentials, issuance is bounded by the subset rule and gated on Mission state. That gated issuance grants task-bound, auditable authority and a possession-independent kill switch at the issuance gate. It grants no per-action control, and outstanding tokens run to their own expiry.¶

Sized deliberately, that expiry is the level's revocation bound. Lifetime-bounded reliance, access-token lifetimes no longer than the deployment's tolerated staleness ([I-D.draft-mcguinness-oauth-mission-status]), gives a quantified cutoff: revocation within one token lifetime, with no Resource Server changes and no status traffic. Expiry closes the temporal bound by the clock alone and observes no revocation, suspension, completion, or containment, so the lifetime must not exceed the tolerated staleness. Revocation latency is a number, not a level: the higher levels add per-action enforcement, parameter binding, and evidence, not a faster clock.¶

Standalone MAS: under a binding without credential-carried authority, Baseline grants governance and audit. No kill switch of any kind exists under such a binding until a freshness surface and runtime enforcement (the next level) arrive, and a deployment states that. The issuance join ([I-D.draft-mcguinness-oauth-mission-issuance-grant]) restores gated issuance at each consuming Authorization Server, and Baseline with it.¶

AAuth: the nearest comparison, stated as capabilities rather than a level, is native approval, exact-byte commitment, active or terminated state, and the ordered mission log. The possession-independent issuance cutoff applies only to person-token issuance and to PS authorization and federated requests whose resource token carries the validated Mission Reference. No Authority Set or subset proof is implied.¶

Proof obligations: the anchored approval and, where credentials are issued, the subset rule.¶

Where the binding gates issuance, a deployment that adds only a freshness surface (Mission Status or introspection with a published staleness bound, [I-D.draft-mcguinness-oauth-mission-status]) gains state-aware reliance: a revocation cutoff within that bound, without per-action enforcement. That is a half-step into the next level, not a level of its own.¶

Verification: illustrative scenarios for this level, each traced to the rule its home document states, are in Appendix E.¶

Runtime-Enforced:

adds a PEP/PDP decision on every consequential action, a trusted state source with a published staleness bound, parameter binding, and runtime evidence ([I-D.draft-mcguinness-mission-runtime] and its AuthZEN profile).¶

Grants: per-action enforcement and bounded revocation (for gated classes, within the staleness bound plus the permit window plus the class's execution bound; for paths no runtime gate reaches, within token lifetime where issuance is gated).¶

Runtime-Enforced is the smallest deployment that turns a Mission from governed issuance into action-time defense. Every normative dependency it needs is a family document intended for the Standards Track. It is a substantial build, not a wedge: a deployment sizes the effort from the runtime profile's conformance section rather than from this level's one-line summary.¶

Proof obligations: PEP-placement completeness and the declared freshness source and bound.¶

Documents: Baseline plus the substrate contract (the kernel that runtime and AuthZEN consume normatively), runtime, its AuthZEN profile, runtime evidence, and a concrete freshness source (Status is the reference choice).¶

Governed Agent (recommended for AI agents):

adds Consent Evidence and the harness, growing with Child Delegation, Expansion, Orchestration, and Discovery (experimental, with Progressive) as needed.¶

Grants: consent-rendering evidence and session-continuity discipline.¶

Documents: Runtime-Enforced plus consent-evidence and the harness.¶

High-Assurance Agent:

adds the guarantees that resist a compromised agent. Two named claims live at this level, each with proof obligations the runtime profile fixes.¶

Agent-compromise-resistant enforcement: mediated (gateway) credential custody, a declared-and-audited path scope, action-bound approval for the high-consequence classes, an active-freshness state source, and approval disclosures rendered by a component isolated from the agent. With these, a compromised agent cannot unilaterally take a high-consequence action for which it does not hold a mediated credential.¶

Trifecta containment: least exposure, the harness taint rule enforced as a mandatory requirement of the harness profile, with pre-consented egress to Approver-named destinations as its one carve-out; and full mediation of the external-communication and external-commitment classes with the egress-channel enumeration. With both halves, an injected agent cannot egress on the strength of untrusted content alone.¶

Both claims are named high bars, never implied by basic adoption. A deployment can bind its Enforcement Scope Statement to execution-environment attestation so a claim is technical rather than organizational ([I-D.draft-mcguinness-mission-runtime], [I-D.draft-mcguinness-mission-harness]).¶

Read in adoption order, each level makes a broader class of agent work defensible to grant.¶

The mapping is informative: the action classes are the runtime profile's ([I-D.draft-mcguinness-mission-runtime]), resource policy remains authoritative for its own objects, and what a level grants varies with the binding. A deployment states its own mediated_action_classes and exclusions in the Mission Deployment Profile (Section 12.6).¶

Every level above Baseline Issuance also carries the cross-cutting obligations its mechanisms imply:¶

• operation-profile normalization where duration or parameter digests are metered ([I-D.draft-mcguinness-mission-metering], [I-D.draft-mcguinness-mission-authzen]);¶

• evidence retention for the audit horizon; and¶

• a registration schedule where audit transparency is run ([I-D.draft-mcguinness-mission-audit]).¶

The evidence mechanisms are accountability, not prevention: they make what was recorded tamper-evident, not what was perceived true or what was never recorded present.¶

11.2. The Reference Architecture

The four levels are also the four reference stacks deployments take, each containing the previous. The stacks are expressed in the OAuth realization. A peer binding realizes the levels per its own document, and peer standing does not imply identical levels or capabilities. The family manifest records each stack's exact membership.¶

• Protocol core (Baseline Issuance): the OAuth binding alone, the standardizable primitive of approved, anchored, state-gated Missions, meeting the Mission Context requirements (Appendix A).¶

• Reference security architecture (Runtime-Enforced): the protocol core plus runtime enforcement, its AuthZEN profile, runtime evidence (the decision and execution objects AuthZEN consumes), and a freshness source (Status is the reference choice). Adoption closure brings in the substrate contract, the normative kernel of runtime and AuthZEN.¶

• Recommended agent architecture (Governed Agent): what a deployment running autonomous AI agents should build.¶

• High-assurance architecture (High-Assurance Agent): the recommended agent architecture plus mediated custody, no unmediated path, action-bound approval, active freshness, and agent-isolated approval rendering.¶

When this document says a Mission is enforced, it means the reference security architecture, and an evaluation should picture it by default. That architecture presumes an authority-bearing binding. Under AAuth, the analogous per-action control is the Person Server's contextual gate on PS-mediated paths ([I-D.draft-mcguinness-mission-aauth]).¶

11.3. Assurance Claims

What a deployment can prove is an axis orthogonal to the levels: named assurance claims, each with a proof obligation an existing profile fixes, listed in the Deployment Profile (Section 12.6) rather than implied by a level:¶

• Approved-record integrity: the committed Intent, authority proposal, and Authority Set reproduce from the retained record alone (the OAuth binding's integrity anchors). The claim covers neither the record's provenance members nor an issuer that substitutes a record and its anchors together; defending against post-approval substitution takes an independently pinned anchor ([I-D.draft-mcguinness-oauth-mission-approved-set-verification]).¶

• Bounded revocation latency, per path and mechanism, naming the paths it covers: for a runtime-gated class, the published staleness bound plus the permit window plus the class's execution bound ([I-D.draft-mcguinness-mission-runtime]); for a lifecycle-gated path, the outstanding credential lifetime; an ungated path has no bound to claim.¶

• Action-time enforcement: PEP coverage for the Enforcement Scope Statement's mediated set, and nothing outside it.¶

• Parameter-bound enforcement: permits bound to concrete parameters for the classes claimed.¶

• Transaction-grade execution: the runtime profile's transaction-assurance tier machinery (single-use permits, leases, outcome reconciliation) for the classes claimed ([I-D.draft-mcguinness-mission-runtime]).¶

• Agent-compromise-resistant enforcement and trifecta containment: the two named High-Assurance claims, as defined in Section 11.1.¶

Two deployments at the same level under different bindings can hold different claims; the MAS modes are the worked case ([I-D.draft-mcguinness-mission-authority-server]). A relying party compares the claims, not the level.¶

11.4. Mission Binding Properties

Whether an operation is bound to a Mission is not one question but three:¶

• attachment provenance: who selected and attached the Mission to this work item?¶

• credential binding: was the acting credential's authority issued and bounded for the Mission?¶

• action binding: does one authenticated permit cover these exact operation inputs?¶

The three dimensions are independent. A native Mission-bound token has strong credential binding with no harness in sight; a trusted harness attributes work items precisely while the credential is an ordinary bearer token; an action-bound permit can exist over either. No single ladder orders them, so the family names the properties directly, as a vector, and a deployment claims the combination each path has.¶

As an index of related properties, the table's entries relate to the three dimensions as follows: mission-reference-selected and work-item-bound to attachment provenance; credential-correlated, credential-mission-bound, presenter-key-bound, and instance-bound to credential binding; and action-bound to action binding. The index establishes no property for any mechanism; the mechanism mapping below states what each mechanism establishes.¶

The mapping join and the Mission Join Assertion are the MAS's ([I-D.draft-mcguinness-mission-authority-server]).¶

The properties are claimed per covered Authorization Server, resource, and action path, never as a product-wide maximum. A mixed estate claims what each path has, and a weaker path never inherits a stronger path's claim from the deployment's name. Where policy requires a property on a path, its absence denies; nothing falls back silently to a weaker binding.¶

Credential-mission-bound is defined by equivalence, not by one artifact. It is the family's authoritative definition of what earns the term "Mission-bound", and it backs the Mission-bound token class (Section 10.2). For the covered path, the credential establishes all of these conditions:¶

1. a trusted issuer authorized to issue for the Mission;¶

2. the canonical (mission.issuer, mission.id) pair identifying the Mission: the baseline Mission reference, and an identity condition only, never a token-carried Authority Set commitment;¶

3. an issued authority projection no broader than the Mission's Authority Set for the target audience;¶

4. the mapped subject, the requesting client_id, and actor or delegation context where applicable;¶

5. bounded lifetime plus active-state issuance and refresh gates; and¶

6. an auditable derivation link from the credential to the Mission's recorded Authority Set: the OAuth binding's Mission Record ([I-D.draft-mcguinness-oauth-mission]), an Issuance Grant jti ([I-D.draft-mcguinness-oauth-mission-issuance-grant]), a cross-domain projection's provenance ([I-D.draft-mcguinness-oauth-mission-cross-domain]), or an equivalently specified artifact. A token-carried Authority Set commitment, where a binding supplies one, strengthens this condition's local verifiability. It is never itself the Mission-identity requirement of condition 2, and its absence does not fail this condition where the issuer-retained record is otherwise reachable.¶

A binding earns the property by one of two routes, never by protocol lineage. Either route must establish all six conditions for the covered path:¶

• Substrate route. A binding that claims substrate conformance supplies its full Mission Substrate Statement ([I-D.draft-mcguinness-mission-substrate], Section "Mission Substrate Statement"): its kernel mappings (among them the Mission Reference, the Controller, the Actor handle and identifier mappings, and the reliance bound) and its capability claims, together with the deployment's applicable issuer trust configuration and the validated credential's own evidence.¶

• Binding route. A binding that claims no substrate conformance establishes the conditions through its own specification and conformance requirements.¶

No fixed split assigns some conditions to the Statement and others to the credential, and a capability label alone never establishes a whole condition. Monotonic Derivation bears on condition 3, and Lifecycle-Gated Authorization on condition 5's active-state gate, but each condition also rests on the mappings, trust, and credential evidence it depends on. The Substrate's Credential-Bound capability can select correlation-only fact semantics ([I-D.draft-mcguinness-mission-substrate]), which by itself evidences neither the derivation-link condition (6) nor the authority-projection condition (3). A Credential-Bound claim alone, whatever semantics it selects, never by itself evidences credential-mission-bound: conditions 1 and 4 also rest on the deployment's issuer trust configuration and the credential's own subject, client_id, and actor claims, which no capability row supplies.¶

The OAuth binding claims no substrate conformance and publishes no Statement, so it takes the binding route: its own Conformance gates establish the conditions, independent of the substrate contract ([I-D.draft-mcguinness-oauth-mission]). Its informative Mapping Assessment explains that mapping and is not a conformance result; OAuth conformance stays governed by the OAuth binding. The OAuth binding's evidence for each condition:¶

Sender constraint is not among the six conditions. Issuance-time key targeting and presentation-time proof are presenter-key-bound, a separate property, so a path that needs possession requires the composition rather than reading possession into the equivalence. The mechanisms match: the OAuth binding recommends sender-constrained tokens, and the generic Issuance Grant leaves cnf optional. Native issuance, the Mission Issuance Grant, and a conforming cross-domain exchange therefore satisfy the equivalence, and supply presenter-key-bound exactly where their confirmation binding is in force.¶

A verified Mission Join Assertion ([I-D.draft-mcguinness-mission-authority-server]) establishes token-specific credential-correlated. For conditions 1, 2, and 4, it authenticates the Mission reference and attests the party join, but trusting the MAS's correlation does not establish that the acting token's issuer issued its authority for that Mission. Conditions 3, 5, and 6 stay unsatisfied by design; that gap is what separates correlation from issuance. The assertion alone establishes neither possession nor instance identity. On the joined path, presenter-key-bound comes from the acting credential and its validated presentation: the MAS mints an assertion only for a token with a cnf key, and the PEP reports that key only after verifying the request's proof of possession, or the certificate only after authenticating it on the request's mutual-TLS connection.¶

Presenter-key-bound is possession and nothing more: the credential names a confirmation key at issuance, and the presenter proves possession at use, with DPoP or mutual TLS. It does not identify which concrete agent process or workload holds the key.¶

Instance-bound requires all of: an authenticated instance identifier; a verified binding from that identity to the confirmation key; current proof of possession; and no key sharing across instances. End-to-end sender constraint alone establishes presenter-key-bound, never this property.¶

Action-bound requires an authenticated permit binding the Mission, the action and resource, and the normalized parameters or complete request projection, under the permit's validity and use controls, in one of two proof forms:¶

• portable permit: an audience-restricted, presenter- or sender-bound artifact, verified where it is presented; the transaction token of [I-D.draft-mcguinness-oauth-mission-transaction-authorization] is the family's discharge; or¶

• channel-bound permit: a permit bound to the mutually authenticated requesting and executing parties and enforced on that same channel, with request cache-key equality, its validity window, and the applicable use controls; the runtime permit of the AuthZEN profile ([I-D.draft-mcguinness-mission-authzen]) discharges it under exactly those conditions.¶

A parameter binding alone makes a response neither form. An evaluation identifier alone is a correlator, never the property; it qualifies only where dereferencing it through an authenticated, audience-bound, freshness- and use-controlled permit store yields one of the two forms above, complete.¶

Deployments claim compositions, and a claimed composition requires every member property:¶

• harness-attributed: work-item-bound, under the attacher and inheritance rules the harness states ([I-D.draft-mcguinness-mission-harness]);¶

• mission-credential-bound: credential-mission-bound plus presenter-key-bound, end to end; instance-bound strengthens the claim where an instance identity exists and is verified; and¶

• runtime-action-bound: Mission binding establishment ([I-D.draft-mcguinness-mission-runtime], Section "Mission Binding Establishment") plus action-bound.¶

The family defines no work-item-attribution composition. Holding work-item-bound and action-bound together does not prove the permitted action came from that work item: concurrent items under one Mission still substitute. The composition becomes definable only when a verified cross-link exists: the action permit, or its authenticated request context, binds the same tamper-resistant work-item identifier the harness recorded. No family carrier supplies that cross-link, so a deployment claims the two properties separately and nothing more.¶

The mechanism mapping is conservative:¶

The property names above are stable identifiers. A claim is a per-path declaration, not prose: each claimed property or composition names the covered issuer, resource, and action-class paths. The Enforcement Scope Statement carries the per-path declarations for the enforcement-adjacent properties, and the Mission Deployment Profile (Section 12.6) composes them. A binding's Statement declares what the binding can supply, which is never itself a deployment claim. An unknown property identifier, an undeclared path, or an unstated property is not claimed, and a consumer treats it as not held; nothing downgrades silently. Schema-level claim identifiers and validation rules remain the Deployment Profile's own future work.¶

Binding properties and the assurance claims above compose rather than repeat. A binding property says what a path establishes, from the Mission a work item is attached to, through the credential, presenter key, and instance, to the authorized action. An assurance claim says what the deployment's enforcement proves. Credential-level and action-level binding likewise compose rather than substitute.¶

11.5. The Containment Matrix

Mission termination is one control in a larger containment surface. Each kill has a different blast radius, and an incident responder needs the whole matrix:¶

The issuer-held containment overlay is defined by [I-D.draft-mcguinness-oauth-mission-containment], and cascade revocation to Child Missions by [I-D.draft-mcguinness-oauth-mission-child-delegation].¶

Mission termination participates in incident response; it does not replace it. Revoking the Mission:¶

• stops issuance at once where the binding gates it;¶

• stops mediated actions within the staleness bound plus the permit window and the class's execution bound (Section 10.4, Section 11.3); and¶

• terminates no process and closes no network path.¶

The converse holds too: killing a workload leaves the Mission active and its authority derivable to a replacement instance unless the Mission is also revoked.¶

Capability kill provides, per action class and consumer, one of two containment properties (named like, but distinct from, the levels of the same name; Section 11.6) or neither, never by the deployment's assurance level ([I-D.draft-mcguinness-oauth-mission-containment], Section "Containment Properties"):¶

• the Baseline property, a new-derivation kill that also propagates to Child Missions justified by the contained entry; and¶

• the Runtime-Enforced property, an action-time kill that also reaches a token issued before the transition.¶

A standalone MAS path with no runtime gate gets neither. Section 11.6 composes the properties by level and binding, with what each leaves running.¶

A deployment's incident runbook names which of these controls exist, who may pull each, and, per action class, which capability-kill property its consumers obtain.¶

11.6. Composed Kill-Switch Reality

The words "Baseline" and "Runtime-Enforced" each have two meanings. In Section 11.1, they name a level a deployment adopts. The containment profile uses the same words for a property a consumer obtains per action class ([I-D.draft-mcguinness-oauth-mission-containment], Section "Containment Properties"). Level and property are not 1:1. A Runtime-Enforced deployment can still provide only the Baseline property for a class its Enforcement Scope Statement leaves lifecycle-gated-only, because the property requires a state-observable substrate per class, not per deployment ([I-D.draft-mcguinness-mission-runtime]). A row can therefore carry a Runtime-Enforced level and a Baseline property together without contradiction.¶

Each row below assumes a deployment that runs the containment profile, at the row's level and under its binding. A level and a binding alone confer neither containment property: containment is an overlay a deployment adopts separately ([I-D.draft-mcguinness-oauth-mission-containment]). In the table, "Stops at commit" is what a contain transition's own state-version commit reaches immediately ([I-D.draft-mcguinness-oauth-mission-containment], Section "The Contain Transition"), and "Runs to its own bound" is the residual the transition does not reach. In the cells, a quoted name is a section of the containment profile ([I-D.draft-mcguinness-oauth-mission-containment]), or of the issuance grant ([I-D.draft-mcguinness-oauth-mission-issuance-grant]) where the cell says so; "runtime profile" is [I-D.draft-mcguinness-mission-runtime].¶

Notes:¶

1. Under the OAuth binding at Baseline Issuance, a cross-domain projection grant already redeemed before the transition runs to its own lifetime ([I-D.draft-mcguinness-oauth-mission-containment], Section "The Materialized-Capability Residual"). Offline attenuation is not accepted at this level (note 6).¶

2. Under the standalone MAS at Baseline Issuance, every action runs to whatever native credential, session, or resource-local bound the resource enforces on its own, if any. It does so until runtime enforcement over a freshness source arrives or the issuance join restores a gate ([I-D.draft-mcguinness-oauth-mission-issuance-grant]).¶

3. Under any binding, a containment-aware state source is full Status or introspection carrying containment_version, or Signals carrying the overlay change.¶

4. Under any binding, where a class is checked only against an active-but-not-containment-aware source, left lifecycle-gated-only, or gated only by fresh derivation, it gets the Baseline property only, regardless of level ([I-D.draft-mcguinness-mission-runtime]). A fresh derivation narrows what it mints and can shorten the residual, but it checks nothing at action time, so it carries the Baseline property, not the Runtime-Enforced one ([I-D.draft-mcguinness-oauth-mission-containment], Section "Containment Properties").¶

5. Under the MAS as estate control plane, a consuming AS either has a Mission-state integration or has none ([I-D.draft-mcguinness-oauth-mission-issuance-grant], Section "Effective Authority Set Projection"). One that has it projects every redemption and refresh through the Mission's current Effective Authority Set, so its issuance excludes contained authority within its published staleness bound. A lifecycle-state check alone, under which a contained Mission still reads active, is not such an integration. One that has none checks no Mission state at redemption, issues no refresh tokens, and relies on the Mission Issuer's minting gate, so a grant minted before the transition can still redeem within its 300-second lifetime.¶

6. Under the OAuth binding with offline attenuation, a consumer accepts a chain only under runtime enforcement, with a fresh active-state check on every presentation regardless of action class ([I-D.draft-mcguinness-oauth-mission-attenuation], Section "The Kill Switch Requires Runtime Enforcement"), so no Baseline-only, lifetime-only path accepts one. Revoking the Mission fails that check, so acceptance of every chain stops within the deployment's declared freshness bound. Containing an entry leaves the Mission active, so the same check does not stop a root minted before the transition from supporting the contained capability: the root keeps minting narrower children for its own lifetime, and its del_max_depth limits how deep they go, not how long they live ([I-D.draft-mcguinness-oauth-mission-containment], Section "The Materialized-Capability Residual"). Claiming the Runtime-Enforced property for attenuated chains needs a containment-aware action-time check (note 3).¶

None of this closes the conforming Baseline residual on a path or for a class no containment-aware action-time gate reaches. For a class a Runtime-Enforced action-time gate reaches instead, a pre-transition credential does not run to its own bound at all. Where the residual does persist, the binding determines the artifact and its cutoff: an ungated standalone-MAS path, for instance, runs to the resource's own bound, if any (note 2), or to none, not to a token lifetime. What changes row to row is which gate, if any, reaches a class before its own bound, and how tight that bound is.¶

11.7. Prevention, Detection, and Residue

Each layer earns a specific property and leaves a specific residue:¶

In each layer, the family commits and checks what a party was shown, decided, or did. It does not make the human attentive, the producer honest, or the unmediated path disappear. Those are the residues the Mission Assurance Levels (Section 11) and the security model make a deployment state rather than assume. The input arm (Section 4.5) carries the same honesty in the other direction.¶

12.1. Entry Ramps by Estate

The estate a deployment already runs, not preference, decides which chokepoint it builds first. The OAuth binding's issuance ramp assumes an Authorization Server that supports pushed authorization requests, rich authorization requests, and JWT access tokens. The standalone ramp assumes none of that and trades it for PEP coverage. By starting condition:¶

Grant redemption and the issuance join are defined by [I-D.draft-mcguinness-oauth-mission-issuance-grant]. Every row shares the record, anchors, and lifecycle, so a ramp is an entry point, not a fork: Missions carry unchanged from any row to the rows a deployment adopts later.¶

12.2. The Issuance-Only Deployment

The entry-ramp table's last row is a deployment in its own right, the issuance-only deployment: the Authorization Server and the Mission-creating client change, and Resource Servers need not be Mission-aware. A delegated token reaches only a Mission-aware Resource Server, so a delegate calling any other resource runs under a Child Mission where child creation is authorized ([I-D.draft-mcguinness-oauth-mission-child-delegation]).¶

The OAuth binding's invariants carry the deployment:¶

• every derived token is a subset of the Authority Set;¶

• no audience receives a grant it would over-grant on; and¶

• no credential outlives the Mission's expiry.¶

The deployment sizes token lifetime to its tolerated staleness. Where the Authorization Server offers introspection, a Resource Server that introspects per request stops honoring a token at its next request once the Mission leaves active, with no Mission-specific code ([I-D.draft-mcguinness-oauth-mission]).¶

The deployment claims approved-record integrity and bounded revocation latency (Section 11.3); Appendix F shows its Deployment Profile. The runtime layer ([I-D.draft-mcguinness-mission-runtime]) joins for the high-consequence classes and for an action class that needs:¶

• per-action evaluation or evidence;¶

• approval bound to a single action; or¶

• a bound the receiving Resource Server cannot enforce.¶

12.3. The Short Mission

One ramp cuts across the rows: the short mission, a Mission whose expires_at sits minutes out, run in records mode with lifetime-bounded reliance. It is a durable approval record with TTL-grade operational cost: audit, anchors, and bounded exposure with no external state-observation surface. The issuer still owns Mission state and the issuance gate (Section 10.4). A deployment can adopt the family this way first, per task, and add state surfaces only where missions grow long enough to need suspend, complete, or kill-now.¶

12.4. The Quarantine Pattern

The quarantine pattern removes a leg of the injection-to-exfiltration chain instead of gating it: no single Mission ever holds untrusted input and an egress path at once.¶

• Work that ingests untrusted content runs under a Mission with no external-communication or external-commitment authority.¶

• Work that communicates externally runs under a separate Mission whose inputs are the quarantined product.¶

• The crossing between them, a human review or a deterministic transformation, is recorded as evidence, under the harness taint policy ([I-D.draft-mcguinness-mission-harness]) and, where claimed, the runtime profile's trifecta containment ([I-D.draft-mcguinness-mission-runtime]).¶

Where untrusted input and egress must stay separate within one Mission, the metering profile's exclusivity control ([I-D.draft-mcguinness-mission-metering]) latches read-and-egress apart under a single approval.¶

The pattern applies the invariants' work-products reading (Section 8) to deployment: a work product is input, not authority, and the producing Mission's authority does not transfer through the artifact by copying, referencing, embedding, or communicating it. The Mission Work Products companion ([I-D.draft-mcguinness-oauth-mission-work-products]) defines the provenance object that attributes an artifact without granting anything.¶

Ingesting a work product adds a conjunctive gate at the receiving Mission's boundary. That gate composes with the three objects' independent gates and does not nest inside them, so Agent identity, Agent Deployment, and Mission stay a gating pipeline, not a containment hierarchy (Section 6.2).¶

12.5. The Standing-Agent Pattern

The standing-agent pattern governs the agent whose work never ends. The agent stands; the authority cycles. The standing thing is a charter: a Mission with a consented authority ceiling and drawdown policy ([I-D.draft-mcguinness-oauth-mission-progressive], experimental). The working thing is the bounded Mission each unit of work draws under the charter, as an in-ceiling successor or a policy-approved Child Mission ([I-D.draft-mcguinness-oauth-mission-child-delegation]). Each unit Mission expires and discharges as its unit completes ([I-D.draft-mcguinness-oauth-mission-status], [I-D.draft-mcguinness-oauth-mission-discharge]).¶

The progressive profile's prohibited set keeps the high-consequence classes on a fresh human approval inside the ceiling. Its Ceiling Review bounds the chain in time with an evidence-rendering renewal.¶

Without the progressive profile, the same pattern runs as ordinary, freshly approved unit Missions with deferred approval ([I-D.draft-mcguinness-oauth-mission-approval]) absorbing the volume: the experimental profile changes the unit economics, not the governance shape.¶

12.6. The Mission Deployment Profile

The Mission Assurance Levels (Section 11) name what to deploy, and the assurance claims (Section 11.3) name what may be proven. A claim is only checkable if a deployment states, concretely, what it enforces and what it leaves outside the boundary.¶

The Mission Deployment Profile is that statement as a system-level artifact: the published composition of the per-layer statements the profiles demand, in one object an auditor, a procurement, or a security review can read. It includes each statement where its profile is run:¶

• the runtime profile's Enforcement Scope Statement;¶

• the harness environment statement;¶

• the MAS mapping contract;¶

• the Resource Server coverage split;¶

• the transparency-service topology and schedule; and¶

• the progressive profile's bounds and ceiling-review cadence.¶

The profile composes existing statements rather than restating them: each fact's owning profile governs its meaning and normative force, and this document fixes no serialization. A machine-readable manifest schema, with stable claim identifiers and validation rules, is deferred family work. Until it exists, the shapes in Appendix F are illustrative, and the per-profile statements are the checkable form.¶

Its distinguishing field is residual_risks: the profile is not credible unless it states, in the same object as its guarantees, what it does not cover.¶

The evidence member carries the deployment's evidence-handling posture beside its guarantees: the field-classification scheme its records use, whether access to Mission evidence is itself audited, and the erasure policy that pairs retention with deletion accountability ([I-D.draft-mcguinness-mission-audit]).¶

The key_custody member declares the custody a deployment states for each signing key it operates, as a list keyed by key and application rather than one row per key class. Each entry states:¶

• its key class, one of the five classes [I-D.draft-mcguinness-mission-security-model] enumerates: issuer signing, evidence signing, agent sender-constraint, mediating-PEP custody, attenuation roots;¶

• the artifact classes or kid selector it covers (the OAuth binding recommends segmenting issuer signing keys by artifact class under distinct kid values within one jwks_uri, [I-D.draft-mcguinness-oauth-mission]);¶

• the holder, and whether the key is exportable;¶

• its generation and signing-use controls;¶

• a reference to its documented compromise-recovery procedure; and¶

• any attestation or verifier reference for that key.¶

software and hsm_or_kms are example holder values naming a mechanism family, not an assurance grade. This document defines neither as a normative custody-grade enum and fixes no validation rule for either. key_custody makes the trusted-base key-custody statement that [I-D.draft-mcguinness-mission-security-model] already requires legible in the Deployment Profile; it does not make that statement checked. Custody assurance stays open until a normative reader or verifier for this declaration exists.¶

Two deployments that both "support Mission" but publish different Deployment Profiles provide different security properties; the profile makes that difference legible.¶

A.1. Context and Intent

• R1: The task an agent pursues is a durable, structured, approved object (oauth-mission; mission-authority-server).¶

• R2: The task and its derived authority are integrity-committed at approval, reproducible from the record alone (oauth-mission).¶

• R3: Task proposals are untrusted input: fields the agent can influence select and narrow what derivation considers, and can request gates the issuer enforces, but never grant or widen authority by their own assertion (oauth-mission; mission-shaping).¶

A.2. Consent and Approval

• R4: The derived authority is disclosed to the Approver before it takes effect, and the approval covers it (oauth-mission).¶

• R5: A single accountable Approver is recorded immutably on the object (oauth-mission).¶

• R6: What was shown at approval is committed and reconstructible by an auditor (oauth-mission-consent-evidence).¶

• R7: Approval can be asynchronous, and any in-review negotiation only narrows (oauth-mission-approval; the experimental oauth-mission-approval-revision).¶

A.3. Lifecycle

• R8: New reliance is gated on task state: only active permits a new governed decision, unrecognized states fail safe, and an already-issued artifact ends at its bounded residual (oauth-mission).¶

• R9: Revocation is independent of credential possession, and state changes propagate by pull or push (oauth-mission; oauth-mission-status; oauth-mission-signals).¶

• R10: A task can be suspended and resumed without being terminated (oauth-mission-status).¶

• R11: Authority widens only through an approved successor: a fresh approval, or policy drawdown within a ceiling a human pre-consented (oauth-mission-expansion; oauth-mission-progressive).¶

• R12: Authority retires per entry when the work an entry served is done (oauth-mission-discharge).¶

A.4. Delegated, Projected, and Enforced Execution

• R13: Derived and delegated authority only narrows (oauth-mission; oauth-mission-attenuation).¶

• R14: Sub-agents receive authority by explicit delegation with lineage, fan-out control, and cascade revocation, never by session ancestry (oauth-mission-child-delegation).¶

• R15: Each consequential action is checked at the point of use, the permit bound to the concrete parameters (mission-runtime; mission-authzen).¶

• R16: When a task stops, governed work stops with it, and in-flight work is classified, then suppressed or cancelled where possible, compensated where authorized, or escalated; irreversible and unknown outcomes remain (mission-harness; mission-orchestration).¶

• R17: Task evidence is tamper-evident and verifiable outside the deployment (mission-audit; mission-mandate).¶

• R18: Two distinct cross-domain properties, never one: a Mission's authority is honorable in another trust domain without widening, through the projection grant (oauth-mission-cross-domain); and a Mission's committed facts are verifiable there without a session with the issuer, granting nothing (mission-mandate).¶

• R19: Delegation history follows authorization continuity, never organizational topology. A Child Mission, an Expansion successor, or any fresh approval starts a new approval basis and actor chain; topology alone neither restarts nor extends one. Representation is profile-specific. Issuer-mediated delegation nests act, and holder-mediated attenuation reconstructs history from per-hop actors. Cross-domain projection carries no upstream act chain, so any destination-domain chain begins locally. Actor identity is attribution and policy input; it does not itself grant or prove authority (oauth-mission; oauth-mission-child-delegation; oauth-mission-expansion; oauth-mission-attenuation; oauth-mission-cross-domain).¶