Network Working Group C. F. Cosme Internet-Draft Triple Optic Systems, Inc. Intended status: Informational 15 November 2026 Expires: 15 May 2027 Delegation and Interaction Traceability for AI Agents: An Attestation and Registry Evidence Model draft-cosme-audit-agent-traceability-00 DRAFT -- NOT SUBMITTED. Independent contribution. draft-cosme-audit-agent-traceability-00 Intended status: Informational Author: Carlos Felipe Cosme, Triple Optic Systems, Inc. Date: 15 November 2026 - Expires: 15 May 2027 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 document is an independent contribution. It does not represent the position of any working group, BoF, or organisation. Abstract AI agents increasingly act through chains of delegation - user to agent, agent to subagent, subagent to tool - in which the party harmed by an action, or asked to trust its result, is rarely the party that operated the agent. This document describes requirements and an evidence model for delegation and interaction traceability in such chains. The model is built on two mechanisms for which measured, independently verifiable evidence exists: (a) an append-only, hash-chained exhibit registry in which each admitted action is recorded exactly once and counters reconcile against registry contents, and (b) per-action attestation published at kernel level with measured publication latencies in the sub-microsecond-to-microsecond range. The document defines a traceability record, states what a verifier can and cannot conclude from a set of such records, and identifies the points at Cosme Expires 15 May 2027 [Page 2] which existing agent-protocol work - identity, authorisation, and transport - needs a common traceability layer. It defines no wire protocol of its own; it states the requirements a traceability mechanism must satisfy so that formats defined elsewhere can be evaluated against them. 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 document may not be modified, and derivative works of it may not be created, except to format it for publication as an RFC or to translate it into languages other than English. 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 2. Terminology 3. Problem Statement 4. Requirements 5. Traceability Record Model 6. Verification and the Limits of Verification 7. Relation to Existing Work 8. Security Considerations 9. IANA Considerations 10. References Appendix A. Measured Evidence Base Author's Address 1. Introduction Cosme Expires 15 May 2027 [Page 3] An AI agent that calls a tool is exercising delegated authority. The user delegates to the agent; the agent may delegate to a subagent; the subagent invokes tools, application programming interfaces, and other agents. When something goes wrong - or when something goes right and a third party is asked to rely on it - the questions are the same: Who authorised this action? Under what delegation? What actually executed? How many times? Can a party that trusts neither the agent nor its operator check the answers? Agent-protocol standardisation is actively addressing identity, authorisation, and inter-agent communication. The "Agent Use of Delegation and Interaction Traceability" (AUDIT) BoF problem space addresses the complementary question: interoperable mechanisms and data models for recording, exchanging, and verifying audit information across user->agent->subagent->tool chains. This document contributes requirements and an evidence model to that problem space, grounded in running, measured implementations rather than in a proposed format alone. Two properties distinguish the model presented here. First, *completeness is a reconciled counter, not an assertion*: an execution counter maintained outside the registry is required to equal the registry's entry count, so an omission is a detectable mismatch rather than a silent gap. Second, *attestation is cheap enough to be per-action*: kernel-level publication of attestation records has been measured at a median of 384-544 nanoseconds per entry, so traceability need not be sampled or batched to be affordable - a design assumption that shapes the requirements in Section 4. This document is Informational. It describes requirements and a record model so that protocol work in the AUDIT and agentproto problem spaces can state which requirements it satisfies. 2. Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here. Principal: The party that originates a delegation (typically a user or Cosme Expires 15 May 2027 [Page 4] an organisation acting through a user). Agent: A software system that takes actions - including tool invocations, message sends, and further delegations - under delegated authority. Subagent: An agent to which an agent delegates a subset of its authority. Tool: A non-agent system whose functions an agent invokes. Delegation event: The act by which a delegator grants a delegate a scoped subset of authority, including the scope, constraints, and expiry of the grant. Action: A single execution performed under delegated authority: one tool invocation, one inference, one message, or one further delegation. Admission: The decision point at which a proposed action is checked against the applicable policy and delegation scope before execution. Traceability record: The evidence artefact produced for an admission or an action, as modelled in Section 5. Registry: An append-only store of traceability records in which each entry is hash-linked to its predecessor, so that modification, deletion, or reordering of any entry is detectable by re-verification of the chain. Verifier: A party that checks traceability records. A verifier is *independent* if it is operated by a party that is neither the agent's operator nor a delegate in the chain being verified. Coverage domain: The explicitly stated set of actions over which a completeness claim is asserted. 3. Problem Statement 3.1. Delegation chains are not reconstructible after the fact Contemporary agent systems log locally, in operator-controlled stores, in formats chosen per product, and frequently only for the outermost agent in a chain. When an action crosses an organisational boundary - the normal case for subagents and hosted tools - no single log contains the chain, and the logs that exist cannot be cross-checked for consistency, omission, or duplication. Reconstruction after an incident is therefore an exercise in trusting the operator whose system is under investigation. Cosme Expires 15 May 2027 [Page 5] 3.2. The three failure shapes Traceability failures in delegation chains take three recurring shapes: 1. Silent omission. An action executes but leaves no record, or its record is later removed. A log that merely contains records cannot, by itself, rebut this failure. 2. Double execution. A delegated action executes more than once - through retry, replay, or a race between delegates - while records show one execution, or none. For actions with external effect (payments, settlements, actuations), double execution is the failure with direct financial and safety consequences. 3. Unattributed action. A record exists but cannot be bound to the delegation that authorised it, so accountability stops at "the system did it." 3.3. Why attestation alone is insufficient An attestation establishes that a particular component produced a particular statement. It does not establish that the set of statements is complete, that no action executed without a statement, or that a statement corresponds to exactly one execution. Conversely, a registry alone establishes ordering and tamper-evidence for what was recorded, but not that recorded statements were true at the point of production. The problem this document addresses is the composition: per-action attestation bound into an append-only registry whose counters reconcile against an independently maintained execution count, so that omission and duplication become detectable events, not unfalsifiable worries. 3.4. Scope This document concerns traceability of delegation and action. It does not define agent identity (addressed by workload-identity work), does not define an authorisation protocol, and does not evaluate the *content* of agent decisions. A traceability record can establish that an authorised action executed exactly once; it cannot establish that the action was wise. 4. Requirements The requirements below are stated for any traceability mechanism claiming to serve delegation chains. They are encoding-independent. REQ-1 (Delegation binding). Every traceability record for an action MUST Cosme Expires 15 May 2027 [Page 6] identify the delegation under which the action was taken, and that delegation MUST itself be traceable to the originating principal through a chain of delegation records, each of which narrows or preserves - never widens - the scope of its parent. REQ-2 (Admission before execution). A mechanism MUST be able to represent the admission decision for an action separately from the action's execution, so that refused actions are as auditable as executed ones. Refusals MUST be recordable in the same registry as admissions. REQ-3 (Exactly-once evidence). A mechanism MUST provide evidence by which a verifier can determine that an admitted action executed exactly once: executions MUST be countable, and a second execution of the same admitted action MUST be detectable as a duplicate rather than indistinguishable from the first. REQ-4 (Completeness by reconciliation). A completeness claim MUST name its coverage domain and MUST be supported by reconciliation of at least two independently maintained counts - for example, an execution counter and a registry entry count - whose equality the verifier can check. A mechanism MUST NOT present possession of a set of records, without such reconciliation, as proof that the set is complete. REQ-5 (Tamper evidence). Registry entries MUST be hash-linked in an append-only structure such that modification, deletion, or reordering of any entry causes verification of the chain to fail. Verification of the chain MUST be performable by an independent verifier given only the registry contents and the chain's public parameters. REQ-6 (Independent verifiability). Verification MUST NOT require trust in the agent or its operator. A verifier that trusts neither MUST be able to check: chain integrity (REQ-5), delegation binding (REQ-1), duplicate detection (REQ-3), and counter reconciliation (REQ-4). REQ-7 (Per-action affordability). Attestation and recording SHOULD be cheap enough, in latency added to the action path, to be applied to every action rather than sampled. Mechanisms SHOULD publish measured publication Cosme Expires 15 May 2027 [Page 7] latencies, including tail latencies, for their recording path, so that "every action" claims are checkable rather than aspirational. REQ-8 (Refusal and failure visibility). A mechanism MUST NOT permit a failing, refused, or unclosed action to be indistinguishable from a successful one. Every admitted action MUST reach a recorded terminal state; an action that neither completes nor records a terminal failure MUST be detectable as unclosed. REQ-9 (Selective disclosure). Records will routinely contain commercially sensitive and personal information. A mechanism SHOULD support disclosure of a record's existence, position in the chain, and verification status without disclosure of its full content, so that verification does not require total exposure. REQ-10 (No format lock-in). The requirements in this section are deliberately encoding-independent. A concrete format satisfies this document by stating, requirement by requirement, how it meets each one - including where it does not. 5. Traceability Record Model 5.1. Record fields A traceability record, at the model level, carries the following fields. A concrete format MAY encode them differently but MUST be able to express each of them (REQ-10). | Field | Meaning | |---|---| | record-id | Globally unique identifier of this record. | | record-type | One of: delegation, admission, execution, refusal, terminal-failure, revocation. | | principal-ref | Reference to the originating principal of the chain. | | delegator-ref / delegate-ref | The parties to this hop of the chain. | | delegation-ref | The record-id of the delegation record authorising this record. | | action-descriptor | A canonical description (or digest) of the action admitted or executed. | | policy-ref | The policy version against which admission was decided. | | sequence | Monotonically increasing position of this record in its registry. | | prev-hash | Hash of the immediately preceding registry entry. | Cosme Expires 15 May 2027 [Page 8] | entry-hash | Hash of this entry, computed over its canonical serialisation including prev-hash. | | attestation | The producing component's attestation over the entry, where the deployment provides one. | | timestamp | Time of recording, with its source stated (see Section 6.3). | | terminal-state | For executions: completed, failed, or unclosed. | 5.2. Registry structure The registry is a single append-only sequence (or a set of sequences with stated merge rules) in which each entry's entry-hash covers the preceding entry's hash. Appending is the only mutation. Two counters are maintained: the registry entry count (derivable from the registry itself) and an execution counter maintained by the execution environment, incremented once per execution. A verification epoch closes when the two counts reconcile and the chain verifies end to end. 5.3. Lifecycle 1. A principal issues a delegation record. 2. For each proposed action, the admission point emits an admission record or a refusal record. 3. On execution, exactly one execution record is emitted, bound to its admission record by delegation-ref and action digest. 4. Each execution reaches a terminal-state. Executions that do not are visible as unclosed at epoch close. 5. Revocation of a delegation is itself a record (revocation); actions admitted after revocation are policy violations visible to the verifier. 5.4. Measured behaviour of a reference implementation The model above is not hypothetical. In a measured run of a registry implementation conforming to this model (November 2026 test campaign; see Appendix A), 240 actions were presented for admission; 186 were admitted and each of the 186 executed exactly once - zero double executions, zero unadmitted executions, zero silent (unrecorded) executions, zero unclosed actions - and at epoch close the independently maintained execution counter (186) equalled the registry entry count for executions (186), with full hash-chain verification returning true. The failures this run was designed to provoke - duplication, silent execution, omission - are precisely REQ-3, REQ-4, and REQ-8 stated as test outcomes. Cosme Expires 15 May 2027 [Page 9] Independently, per-action attestation publication at kernel level was measured over 756 published entries at a median (p50) of 384-544 nanoseconds and a 99th percentile (p99) of 1.8-2.5 microseconds per entry, with an independent verifier passing all 16 of its checks - evidence for REQ-7 that per-action recording need not be sampled. 6. Verification and the Limits of Verification A traceability mechanism is only useful if its claims are stated as narrowly as they are earned. This section states the limits explicitly; a format or verifier conforming to this document MUST preserve these distinctions in anything it reports. 6.1. What verification establishes Given a registry that verifies under REQ-5 and reconciles under REQ-4, a verifier can establish: that the records presented are the records that were appended, in the order appended; that no admitted action in the coverage domain executed more than once without detection; that every execution in the coverage domain reached a recorded terminal state; and that each execution was bound to a delegation chain terminating at a stated principal. 6.2. What verification does not establish Verification does not establish that recorded statements about the *content* of an action are true - only that they were recorded, attested by the stated component, and not altered afterwards. It does not establish completeness outside the stated coverage domain. It does not establish the independence of a signer from the operator merely because distinct keys were used: independence is a property of control, not of key separation, and a verifier report MUST NOT claim independence it has not established on control grounds. 6.3. Time A timestamp in a record establishes, at most, that the record existed by the time it was anchored in a structure the verifier trusts (for example, its position in a hash chain whose head is externally witnessed). Timestamps from the producing component's local clock establish ordering within that component's records and nothing stronger. Verifier reports MUST distinguish Cosme Expires 15 May 2027 [Page 10] anchored time from local-clock time. 6.4. Reporting unauthenticated content Where a record contains fields not covered by an attestation or not anchored in the chain, a verifier MUST report those fields as unauthenticated. A verification success for the chain MUST NOT be presented as authentication of every field of every record in it. 7. Relation to Existing Work Agent communication protocols (agentproto problem space) define how agents discover, address, and communicate with one another. This document is the audit companion to that work: whatever protocol carries the interaction, the traceability requirements here apply to the delegation and action records the interaction generates. Workload identity and authorisation (WIMSE and related work) establish who a workload is and what it may do. Traceability consumes those results: delegator-ref, delegate-ref, and policy-ref are references into identity and authorisation systems, not replacements for them. Transparency and attestation building blocks. Hash-chained append-only logs, remote attestation evidence, and verifiable data structures are established IETF building blocks. This document composes them for the delegation-chain case and adds the two requirements the composition needs: exactly-once evidence (REQ-3) and completeness by counter reconciliation (REQ-4). Policy enforcement measurements in a related implementation - zero-copy GPU-to-kernel BPF policy enforcement at 111,792 policy updates per second with p99 latency of 8.9 microseconds - indicate that admission- time policy checks (REQ-2) are likewise affordable on the action path. Other audit-format drafts. Several individual drafts in the AUDIT problem space define record formats and architectures. This document is deliberately format-free so that those drafts can be evaluated against common requirements, and so that their limits (Section 6) can be stated once and inherited by reference. 8. Security Considerations Registry withholding. A registry operator can withhold records rather than alter them. Hash-chaining detects alteration, not withholding. Deployments that need withholding resistance SHOULD anchor registry Cosme Expires 15 May 2027 [Page 11] heads with independent witnesses and MUST state, in any completeness claim, the witnessing arrangements the claim relies on (REQ-4's coverage domain includes its witnessing basis). Counter capture. Completeness by reconciliation (REQ-4) fails if both counters are controlled by the same party that benefits from an omission. The execution counter MUST be maintained by a component outside the control of the party whose actions are being recorded, or its integrity MUST be attested and anchored equivalently. Key compromise and attestation forgery. A compromised attestation key lets an attacker mint records indistinguishable from genuine ones until the compromise is detected. Deployments MUST support revocation records, key rotation, and epoch boundaries at rotation, so that post-compromise records can be quarantined by epoch. Replay. Exactly-once evidence (REQ-3) depends on action digests being bound to admission records; a format that permits an execution record to be replayed under a new admission defeats it. Formats MUST bind execution to admission by a value an attacker cannot re-mint. Privacy. Traceability records are, by design, a detailed account of who did what. They are sensitive. REQ-9 (selective disclosure) is a security requirement, not a convenience: a mechanism that requires full disclosure to verify will either leak or not be used. Retention, access control, and data minimisation for registries are deployment obligations this document flags but does not specify. Denial of service via the record path. If recording is on the action path, an attacker who can stall recording can stall action. Recording paths SHOULD fail closed for actions whose policy requires a record, and SHOULD publish their latency behaviour (REQ-7) so that the cost of the control is visible. 9. IANA Considerations This document has no IANA actions. A future format document adopting this model may request a registry for record-type values; the initial values in Cosme Expires 15 May 2027 [Page 12] Section 5.1 are offered as the starting content for such a registry if the problem space wants one. 10. References 10.1. Normative References - [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, March 1997. - [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, May 2017. 10.2. Informative References - [RFC9334] Birkholz, H., et al., "Remote ATtestation procedureS (RATS) Architecture", RFC 9334, January 2023. - [RFC9162] Laurie, B., Messeri, E., and R. Stradling, "Certificate Transparency Version 2.0", RFC 9162, December 2021. (Informative: an existing append-only, hash-structured transparency log design.) - AUDIT BoF, "Agent Use of Delegation and Interaction Traceability", BoF problem statement and proposed charter discussion, IETF 127, November 2026 (mailing list: audit@ietf.org). - IETF agentproto working group charter, chartered October 2026 (agent communication protocol building blocks). Appendix A. Measured Evidence Base All measurements below are from the author's implementations, are stated exactly as measured, and are the sole quantitative evidence relied on in this document. Existence-level description only; no implementation detail beyond architecture is disclosed. A.1. Registry exactly-once run (N-3 Run 3). 240/240 admissions processed; 186 executions, each exactly once; 0 double executions; 0 unadmitted executions; 0 silent (unrecorded) executions; 0 unclosed actions; execution counter 186 == registry execution count 186; hash-chain verification: TRUE. A.2. Per-token kernel-map attestation publication (measured 5 November 2026). 756 entries published during live serving; publication latency p50 384-544 ns, p99 1.8-2.5 us; independent verifier: PASS on all 16 checks. A.3. Kernel policy enforcement (F0 platform measurement). Zero-copy GPU->kernel BPF policy enforcement: 111,792 policy updates/sec, p99 8.9 us. Cosme Expires 15 May 2027 [Page 13] Author's Address Carlos Felipe Cosme Triple Optic Systems, Inc. Email: carlosfcosme@thetripleoptic.com *Triple Optic Systems, Inc. - formation pending at the time of drafting (5 November 2026). The author contributes in an individual capacity; this document claims no organisational membership or endorsement.*