The Agent Action Decision Protocol (AADP): Per-Action Authorization for AI Agents

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶

This Internet-Draft will expire on 18 February 2027.¶

Copyright Notice

Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶

This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶

Table of Contents

1. Introduction

AI agents increasingly perform consequential actions: moving money, modifying data, sending communications, executing code, and provisioning infrastructure. The frameworks that host these agents typically decide what an agent can do by controlling which tools are exposed to it. Recent analysis has shown that this conflates two distinct questions: capability (whether an agent can invoke a tool) and authorization (whether this call, with these argument values, should be performed now) [SCOPEGATE]. A tool that is legitimately exposed for one purpose can be invoked for another; an agent authorized to make one payment can attempt a second; and a generic capability such as an HTTP client can reach effects its name does not declare.¶

Parallel work standardizes agent identity and credential flows [I-D.klrc-aiagent-auth] [I-D.ni-wimse-ai-agent-identity] and hardens the authorization of protocol connections [MCP]. That work answers who the agent is and what it may reach; the WIMSE applicability work states explicitly that whether an agent inherits its user's permissions is out of its scope. It deliberately leaves per-action concerns -- budget enforcement, obligation tracking, approval lifecycle -- as deployment-specific. AADP specifies exactly that remainder: a small, implementable contract for per-action authorization decisions, designed to compose with identity-layer standards rather than replace them.¶

Closer to this document, [I-D.liu-agent-operation-authorization] authorizes individual operations delegated from a human principal, carrying policy as Rego, signed confirmation evidence, and a delegation chain across multiple agents. It addresses a different remainder: consent and delegation lineage. It does not constrain cumulative spend, and the approval it records is a binary confirmation rather than a lifecycle with expiry and re-evaluation. AADP is complementary to it: the state that must persist between calls -- budgets shared across differently-named tools, reservations that outlive a decision, approvals that can expire -- is what this document specifies, and it cannot be expressed by evaluating one call at a time however the policy is written.¶

The Policy Decision Point / Policy Enforcement Point split, and the concept of obligations attached to a permission, originate in XACML [XACML]. AADP inherits that vocabulary deliberately and differs in what it standardizes: agent-action semantics that stateless policy evaluation cannot express -- stateful budget reservation shared across differently-named tools, autonomy tiers with reversibility as a precondition, a two-phase exchange separated by a durability boundary, and typed evidence sufficient to re-derive every verdict.¶

2. Roles

Policy Decision Point (PDP):

Evaluates a proposed action and returns a verdict. Owns policy, budget state, approval state, and the evidence record. There is exactly one logical PDP per governed domain.¶

Policy Enforcement Point (PEP):

Sits wherever actions actually happen -- an agent-framework tool wrapper, a protocol proxy, an API gateway filter, a container or micro-VM supervisor, a workflow engine. A PEP MUST NOT perform a governed action without a permit, and MUST report the outcome of every permit it receives.¶

A deployment MAY have many PEPs consulting one PDP. Every PEP is bound by the same contract.¶

3. Protocol Model

4. The Two-Phase Exchange

AADP is a two-phase protocol. The phases are separated by a durability boundary because the act between them may be slow, remote, or fatal.¶

 PEP                              PDP
  |  decide(request)               |
  |------------------------------->|  evaluate, reserve budget,
  |                                |  record intent, COMMIT
  |<-------------------------------|
  |  permit(permit_id, obligations)|
  |                                |
 [ PEP performs the action ]       |
  |                                |
  |  report(permit_id, outcome)    |
  |------------------------------->|  record linked result, COMMIT
  |<-------------------------------|

A PEP that receives a permit MUST send exactly one report, whatever happened -- success, failure, timeout, or refusal to act. A PDP MUST treat a permit with no report as an unresolved intent and MUST NOT release its budget reservation implicitly.¶

An unresolved intent therefore holds budget that no action ever spent, and a caller that requests permits and never reports can exhaust a budget at the cost of one decide call per permit, leaving no effect at the sink to detect. Reservations MUST therefore be reclaimable explicitly rather than only implicitly: a PDP that issues a permit carrying "execute_within" (Section 6) MUST, once that deadline has passed with no report, append an expiry entry to the evidence record and release that permit's reservation. The release is an audited event, not a silent timeout, and the permit is void from that point. A PDP SHOULD attach "execute_within" to every permit that reserves budget; one that does not MUST document how reservations are otherwise reclaimed.¶

Combining the two phases into a single call is not conformant: a decision point that waits for the act holds state across an operation it does not control.¶

5. Messages

All messages are JSON objects [RFC8259]. Timestamps are [RFC3339] in UTC. Monetary values are decimal strings (for example, "12.50"), never floating-point numbers.¶

{
  "protocol": "aadp/0.1",
  "request_id": "0a5f2c3e-1b7e-4f2a-9c1d-6b0f9e2a77c1",
  "action_type": "payments.transfer",
  "params": { "payee": "acme-gmbh", "amount_eur": "40.00" },
  "source": "llm",
  "rationale": "user asked to settle invoice 8841",
  "session_id": "sess-91f3",
  "cost_eur": "40.00",
  "created_at": "2026-08-17T09:14:02Z",
  "parent_id": null
}

{
  "protocol": "aadp/0.1",
  "request_id": "0a5f2c3e-1b7e-4f2a-9c1d-6b0f9e2a77c1",
  "verdict": "permit",
  "reason": "passed",
  "detail": "",
  "nominal_tier": 1,
  "effective_tier": 1,
  "permit_id": "pmt-h8Qk2mAoT7xY",
  "obligations": [
    { "type": "report_result", "required": true },
    { "type": "undo_available_until",
      "value": "2026-08-17T09:29:02Z" }
  ],
  "approval_id": null,
  "evidence_id": "aud-88213"
}

{
  "protocol": "aadp/0.1",
  "permit_id": "pmt-h8Qk2mAoT7xY",
  "outcome": "success",
  "payload": { "confirmation": "TX-77120" },
  "error": null,
  "completed_at": "2026-08-17T09:14:03Z"
}

{ "protocol": "aadp/0.1", "accepted": true,
  "evidence_id": "aud-88214" }

6. Obligations

An obligation is a condition attached to a permit that the PEP MUST satisfy for the permission to be valid. Obligations are typed and extensible; they are the protocol's extension point for enforcement capabilities the PDP itself cannot provide.¶

Unknown obligations fail closed. A PEP that receives an obligation whose type it does not recognize MUST NOT perform the action, and MUST report "not_attempted" with an error naming the unsupported type. A PEP MUST NOT silently ignore an obligation: a policy author has to be able to assume that an issued obligation was either honoured or visibly refused.¶

Where an obligation admits evidence of its discharge (for example, the identity of the isolation boundary used), the PEP SHOULD include that evidence in its report payload, so that the discharge is recorded alongside the decision.¶

Vendor-specific obligations MUST use a vendor prefix as in Section 3.3.¶

7. Idempotency and Replay

"request_id" is the idempotency key. A PDP that receives a decide request for a request_id it has already decided MUST return the original verdict with "verdict" and "reason" unchanged, and MUST NOT re-reserve any budget. Implementations MAY additionally set "replayed": true.¶

A "permit_id" and an "approval_id" are capability identifiers: possession of one is sufficient to report an outcome for a permit, or to resume an approved action. They MUST therefore be unpredictable -- infeasible to guess from other issued identifiers -- so that a party holding only a decide credential cannot forge a result, or a resumption, for an exchange it did not conduct. A sequential or otherwise enumerable identifier does not satisfy this requirement. Single-use bounds replay of a known identifier; unpredictability bounds forging of an unknown one. Where a deployment cannot guarantee unpredictable identifiers, the report and resume channels MUST instead bind each identifier to the authenticated identity that received it, and reject any report or resumption presented under a different identity.¶

A "permit_id" is single-use. A second report for the same permit_id MUST be rejected with "accepted": false.¶

Resuming an approved action (Section 8) is a new decision with a new request_id; reusing the proposed request's id would return the "propose" verdict forever.¶

8. Approvals

When the verdict is "propose", the PDP creates an approval record with a deadline and returns "approval_id". Approval state transitions are pending to approved, denied, or expired.¶

A conformant PDP MUST: transition pending to approved atomically, rejecting a second approval and any approval after the deadline; record who approved and when; and re-evaluate the action on resumption. An approval authorizes the action, not the outcome: a kill switch engaged between approval and resumption MUST still deny.¶

A granted approval is single-use: the PDP MUST mark it consumed when a resumed request is decided against it, so that one human decision cannot authorize repeated actions.¶

Resumption may be driven from either side, and both forms are conformant provided the guarantees above hold. In PEP-driven resumption, the PEP issues a fresh decide request carrying "approval_ref". In PDP-driven resumption, granting the approval is itself the resumption: the PDP re-evaluates the recorded request under a fresh request_id and returns a decide response. A PDP MUST document which form it implements, and MUST NOT let either form bypass re-evaluation.¶

Approval transport -- who is notified, and how -- is out of scope.¶

9. Evaluation Invariants

This section is the normative core. Two conformant PDPs with the same policy MUST reach the same verdict for the same request. Implementations MAY optimize freely -- caching, compiling, reordering pure checks -- provided these observable orderings hold.¶

1. The kill switch is evaluated first, before policy lookup. An engaged kill switch MUST NOT be overridable by a policy entry.¶
2. Unknown actions are denied. Absence of policy is a denial with reason "default_deny", never a permission.¶
3. Irreversibility escalates. An action at tier "auto" with no registered means of reversal MUST NOT auto-execute; it escalates to "confirm" with reason "no_compensating_command".¶
4. Bounds are validated before proposal. A human MUST NOT be asked to approve an action whose parameters already violate policy.¶
5. Rehearsals do not spend. If an action is in dry-run, the PDP MUST NOT reserve any budget.¶
6. Budget reservation is atomic and all-or-nothing. Where an action draws on several budgets, all MUST be checked before any is reserved, and a failure on any MUST leave every counter untouched. Concurrent requests MUST NOT be able to over-reserve a budget.¶
7. Escalation is recorded, not hidden. When effective tier exceeds nominal tier, the reason MUST identify the escalating check.¶
8. Reversal is governed. An undo is itself an action subject to the full pipeline. There is no privileged path.¶
9. Intent precedes action. The evidence record of a permit MUST be durable before the permit is returned. A crash between phases MUST leave a recoverable, detectable state: an intent with no result.¶

10. Evidence

A conformant PDP MUST maintain an append-only record in which each entry carries at minimum: the request, the verdict, the reason code, both tiers, a timestamp, and -- for results -- a link to the intent it resolves.¶

Entries MUST NOT be updated or deleted. Corrections are appended, never applied in place. Implementations SHOULD enforce this structurally (storage-level constraints), not only by convention.¶

The durability an intent must have before its permit is returned (invariant 9) means, at minimum, survival of the failure of the PDP process: a permit that was returned MUST be recoverable after the deciding process restarts. Whether the record also survives the failure of the host -- power loss, operating-system crash -- is a deployment property a PDP MUST document, since it determines whether an intent can be lost in a window the enforcement point cannot observe. Returning a permit whose intent is not durable to the documented level is non-conformant.¶

The record MUST be sufficient to re-derive every verdict it contains, given the policy version in force at the time. A PDP MUST therefore identify, on every entry, the policy that produced it, and MUST retain that policy for as long as the entries referring to it are retained. Identifying the policy by a content digest of its normative content is RECOMMENDED, so that an unchanged policy keeps one identity and a reverted change is recognisable as a return to an earlier one rather than as a third state.¶

This requirement exists because policy is usually stored mutably. A decision point that overwrites policy in place, and records only the verdict, has an evidence record that cannot be checked: it cannot show that a permit issued last week was correct under last week's rules, and an operator who loosens policy, acts, and restores it leaves nothing behind. Retaining superseded policy is what makes the rest of this section meaningful.¶

11. Failure Semantics

If a PEP cannot reach the PDP, its behavior is a deployment policy that MUST be configured explicitly, not defaulted silently:¶

fail_closed (RECOMMENDED):

perform no governed action.¶

fail_static:

permit only actions on a locally cached list of statically decidable, unbudgeted, tier-0/1 permissions, and queue evidence for later delivery.¶

fail_open:

perform the action. Conformant implementations MUST NOT make this the default and SHOULD require it to be named in configuration.¶

A PEP operating under fail_static MUST deliver queued evidence when the PDP becomes reachable, and MUST NOT locally permit any action that would require a budget, an approval, or an obligation.¶

These modes govern a PDP that could not be reached. A failure that indicates the request itself was not well-formed -- a transport or parse failure of the kind described in Section 5.1 and Section 12.1 -- MUST be treated as fail_closed regardless of the configured mode: such a request MUST NOT be performed under fail_open or fail_static. A malformed request must not become a permitted action by way of a failure policy, just as it must not by way of the decision function.¶

12. Transport Bindings

The core protocol is transport-agnostic. Two bindings are defined.¶

13. Conformance

A conformant PDP returns a verdict for every request it accepts, including one it cannot parse (Section 5.1). It implements the messages of Section 5, the invariants of Section 9, the idempotency rules of Section 7, the approval semantics of Section 8, and the evidence requirements of Section 10.¶

A conformant PEP never performs a governed action without a permit; sends exactly one report per permit; fails closed on unknown obligations; and has an explicitly configured unreachability behavior.¶

A minimal PEP MAY implement only decide and report and refuse any verdict other than "permit" and "deny". It MUST then treat "propose" and "dry_run" as refusals to act and report "not_attempted".¶

14. Versioning

The "protocol" field carries "aadp/MAJOR.MINOR". Minor versions add optional fields, reason codes, and obligation types; a recipient MUST ignore unknown fields. Unknown obligations and unknown verdicts are the exception: both MUST fail closed. Major versions may change required semantics.¶

15. IANA Considerations

This document requests the creation of an "Agent Action Decision Protocol (AADP)" registry group with three registries, each operating under the Specification Required policy [RFC8126]. Names beginning with "x-" are reserved for vendor use and are not registered.¶

Registration requests for all three registries use the same template: Name (the wire identifier), Description (one sentence), Change Controller, and Reference (the specification defining the semantics). For the Obligation Types registry, two further fields are required: Value Syntax, and Discharge Evidence (what a PEP MUST supply in a report to demonstrate the obligation was met). The designated expert should confirm that the semantics are implementable by a PEP without access to PDP-internal state, and that no existing entry already expresses them.¶

16. Security Considerations

Parameters are untrusted input, commonly authored by a language model that may itself be under prompt-injection influence. PDPs MUST validate parameters against declared bounds and SHOULD reject unknown parameters rather than pass them through. The "rationale" field is model-authored text in the common case; it is evidence, never input to the decision.¶

AADP mitigates a class of confused-deputy failures [SCOPEGATE] by evaluating argument values per call rather than gating tool exposure. It does not defend against a PEP that can act without consulting the PDP: a capability reachable outside the governed perimeter is outside this protocol's ability to detect. Deployments SHOULD verify that no unsupervised path to a governed capability exists.¶

The interval between a permit and its execution is a time-of-check/time-of-use window. The "execute_within" obligation bounds it; deployments with strict requirements SHOULD issue it. A revoked or expired permit discovered at execution time is reported "not_attempted".¶

The approval queue is a denial-of-service surface: an agent that can flood proposals can bury a human approver. Rate caps on proposal-generating action types, and approval TTLs, bound the queue.¶

Budget reservation is a second denial-of-service surface, and a quieter one. A caller that requests permits and never reports consumes the reserved budget without performing any action, so nothing appears at the sink and nothing fails; legitimate work is simply denied for the rest of the window. The explicit reclamation required in Section 4 bounds the exposure to the "execute_within" deadline, which is why that obligation is RECOMMENDED on every budgeted permit rather than reserved for time-critical actions.¶

The evidence record may contain sensitive parameter values. Implementations SHOULD support field-level redaction at write time and MUST NOT achieve redaction by mutating existing entries. Access to the evidence record SHOULD be governed no less strictly than the actions it records.¶

Monetary values as decimal strings avoid floating-point rounding as an attack surface on budget arithmetic.¶

Appendix A. Worked Exchange (Tier 3)

-> decide  { request_id: A, action_type: "payments.transfer",
             cost_eur: "4000.00" }
<- propose { reason: "effect_floor", nominal_tier: 2,
             effective_tier: 3, approval_id: "apr-9Xt4Bs7Q" }

   [ human approves apr-77 ]

-> decide  { request_id: B, ..., approval_ref: "apr-9Xt4Bs7Q" }
<- permit  { permit_id: "pmt-Rc8Hn2Vw5Lp3",
             obligations: [ {type:"report_result"},
               {type:"execute_within",
                value:"2026-08-17T09:20:00Z"} ] }

   [ PEP performs the transfer ]

-> report  { permit_id: "pmt-Rc8Hn2Vw5Lp3", outcome: "success" }
<- accepted { evidence_id: "aud-88221" }

Appendix B. Implementation Status

This section records the status of known implementations, per the practice described in [RFC7942]; it is to be removed before publication as an RFC.¶

onedoor (Apache-2.0, Python, version 0.3.4 on PyPI at the time of writing) is the reference implementation of the PDP, with a test suite covering the invariants of Section 9 and running on Linux and Windows. Its decision service exposes the HTTP operations of Section 12.1.¶

Four enforcement points have been written against it, at differing levels of maturity. Two are packaged with the implementation: an in-process library binding, and a Model Context Protocol stdio proxy. Two are experimental examples rather than supported adapters: a LiteLLM gateway guardrail and a LangGraph tool wrapper with interrupt-based approval flow. The LiteLLM example is known to deviate from Section 4 in that it reports on issue of the permit rather than after the act; it is therefore not conformant as written, and is included here as evidence that the gateway hook point is viable, not as a conformant PEP.¶

onedoor records policy provenance as required by Section 10: the policy set is content-hashed, retained in an append-only table, and the hash is stamped on every evidence entry. This was added after the requirement was written. Its cost, from interleaved A/B runs on one host, is roughly 4 per cent of a decide-and-report cycle -- around 0.02 ms on a cycle of about 0.49 ms. Single-run figures at this scale are not distinguishable from run-to-run variation, which is itself of that order, so the figure is offered as an order of magnitude and not as a benchmark.¶

Its evidence writes are split along the durability boundary this document draws: the intent row is always written synchronously before the permit is returned, while result rows may optionally be batched. Batching results costs nothing this document requires -- a lost result leaves an intent with no result, which is the recoverable state Section 9 already mandates -- and on one host it raised throughput from roughly 1,940 to 2,840 decide-and-report cycles per second, saturating at a batch of about 16. Exactly-once reporting is preserved by checking the pending batch as well as the record.¶

onedoor does not yet implement the explicit reservation reclamation required by Section 4: a permit that is never reported holds its budget until the accounting window rolls over. The requirement was added after probing the implementation, and the implementation has not caught up.¶

Two further requirements of this revision are not yet met by the reference implementation, and are recorded here rather than hidden. Its permit identifiers are sequential rather than unpredictable (Section 7); a deployment relying on it therefore requires the identity-binding alternative that section permits. Its evidence is durable to process-crash but not to host-failure level (Section 10), which it documents rather than strengthens. A conformant unpredictable-identifier scheme and a host-durable evidence mode are planned but unshipped.¶

onedoor implements PDP-driven resumption (Section 8): its approve operation re-evaluates the recorded request under a fresh request identifier and returns a decide response. It does not currently accept "approval_ref" on a decide request; PEP-driven resumption is therefore specified but not yet exercised by an implementation.¶

Appendix C. Acknowledgments

The PDP/PEP decomposition and the obligation concept derive from XACML. The capability-versus-authorization framing was independently articulated by David Mellafe Zuvic in the analysis cited as [SCOPEGATE], whose limitations section names stateful side channels as future work; this document treats that remainder as part of its scope.¶

Author's Address