]> Independent

US enrique@somoza.co

AI agents trust capability delegation provenance handshake DNS DNSSEC SCITT RATS This document defines the Agent Trust Negotiation (ATN) protocol. ATN sits above existing DNS-based agent discovery mechanisms and answers questions that discovery alone cannot: what is the agent permitted to do, under whose authority, with what provenance, and how do two agents reach a verifiable working agreement. ATN binds four artifacts to a DNS-discovered agent identity:

1. A Capability Manifest expressing what the agent is willing to do, under what conditions, and with what constraints.
2. A Delegation Chain proving authorized principal scope and revocation paths.
3. A Provenance Attestation proving build-time and runtime integrity.
4. A Session Receipt produced at handshake completion, recording the negotiated scope.

ATN defines the handshake state machine that consumes these artifacts to produce a mutually verified, scope-bounded agent session, and a session receipt suitable for transparency log inclusion. ATN is transport-independent above the DNS layer and compatible with existing discovery proposals including AID , DNS-AID , and ANS v2 .

Introduction Recent IETF work has converged on DNS-based discovery for AI agents. Multiple drafts now define how to locate an agent's endpoint and verify its cryptographic identity given a domain name. These efforts answer two questions: where is the agent and who is the agent. They do not answer the questions that arise immediately after:

• What is the agent permitted to do, and under what constraints?
• Under whose authority is the agent operating?
• What is the provenance of the agent's software, model, and configuration?
• How do two agents reach a mutually verifiable working agreement before any task begins?
• What evidence remains after the session ends, and how is it audited?

This document defines the Agent Trust Negotiation (ATN) protocol to address these gaps. ATN is the layer above discovery. It assumes the discovery layer has located an endpoint and authenticated an identity through one of the existing mechanisms. ATN then negotiates the trust relationship. The design has four parts:

1. Four structured artifacts (Capability Manifest, Delegation Chain, Provenance Attestation, Session Receipt) that can be pointed to from any DNS-based discovery record.
2. A handshake protocol over HTTPS that exchanges and verifies the artifacts between two agents.
3. A capability intersection algorithm that yields the negotiated session scope.
4. A receipt format suitable for SCITT-style transparency log inclusion.

ATN deliberately does not introduce a new discovery mechanism. It builds on existing work and adds the negotiation layer that the agent ecosystem currently lacks.

Scope In scope:

• Format and semantics of the four ATN artifacts.
• The Agent Trust Handshake (ATH) state machine.
• Capability intersection algebra.
• DNS bindings to point from a discovery record to ATN artifacts.
• IANA registrations for ATN-specific identifiers.

Out of scope:

• Discovery of the agent itself (use , , , or equivalent).
• Endpoint authentication (delegated to the discovery layer).
• Application protocol (delegated to MCP, A2A, OpenAPI, or equivalent).
• In-organization service-to-service identity (see ).
• Reputation distribution.
• Behavioral attestation.

Non-Goals

• ATN does NOT define a new DNS record type.
• ATN does NOT replace any existing discovery proposal.
• ATN does NOT introduce a centralized trust authority.
• ATN does NOT enforce policy. It produces verifiable evidence of declared and negotiated policy that downstream systems can audit.

Choice of Venue ATN's choice of IETF as the venue follows from its dependencies. The discovery substrate is DNS and SVCB , both IETF standards. The attestation architecture is RATS . The transparency log is SCITT. The signature format is JWS . The mutual authentication transport is TLS . The authorization frameworks ATN composes with are OAuth and GNAP . ATN is the composition of IETF primitives into a peer-to-peer agent trust handshake. Standardizing the composition in any other venue would fork the primitive stack and create alignment problems. The natural home for the composition is the same venue that standardized the primitives.

Conventions and Definitions 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 when, and only when, they appear in all capitals, as shown here.

Terminology

Agent

An automated software entity that initiates network interactions on behalf of, or with delegated authority from, a principal, without per-interaction human authorization.

Initiator

The agent that opens the trust handshake.

Responder

The agent that answers the trust handshake.

Capability Manifest

A signed JSON document declaring what an agent is willing to do, under what conditions, and with what constraints.

Delegation Chain

A signed sequence of authorization statements binding the agent to a principal, with explicit scope, time bounds, and revocation pointers.

Provenance Attestation

A signed document proving build-time and runtime characteristics of the agent.

Session Receipt

A signed record produced at handshake completion, summarizing the negotiated scope and identities, suitable for transparency log inclusion.

Negotiated Scope

The intersection of two agents' capability sets, time windows, authorization scopes, and operational constraints, as determined by the handshake.

Architecture Overview ATN defines four roles:

1. Agent Operator: publishes the DNS discovery record and the ATN artifacts.
2. Agent: the runtime entity, in either Initiator or Responder mode at any moment.
3. Principal: the human, organization, or other agent that authorized the agent through the Delegation Chain.
4. Witness: an optional third party that can pre-verify ATN artifacts and publish endorsements (e.g., via SCITT ).

ATN layers above discovery: The handshake order is:

1. Discovery: Initiator resolves the Responder's discovery record via DNS.
2. Artifact fetch: Initiator retrieves the four ATN artifacts referenced by the discovery record.
3. Handshake: Initiator opens the ATN handshake. Both parties exchange artifacts and verify signatures.
4. Negotiation: Both parties compute the capability intersection.
5. Receipt: Both parties produce a signed Session Receipt covering the negotiated scope.
6. Session: Application traffic proceeds within the negotiated scope.

DNS Binding ATN does not define its own DNS records. Instead, it specifies how to point to ATN artifacts from existing discovery records.

Binding Tags for TXT-Based Discovery The following tags MAY be added to TXT-based discovery records (compatible with the format and the format defined in companion documents):

Tag	Description
atn	Absolute HTTPS URI of an ATN Index Document.
atn-digest	Hex SHA-256 digest of the Index Document.

The ATN Index Document is a JSON file pointing to the four artifacts. Indirection through an Index Document keeps the DNS record small.

Example (binding ATN to an AID-style record)

Binding ParamKeys for SVCB-Based Discovery For SVCB-based discovery , the following ParamKeys are registered (see ):

ParamKey	Format
atn	URI
atn-digest	hex-string

Example (binding ATN to a DNS-AID-style SVCB record)

ATN Index Document The Index Document at the atn URI is a JSON object with the following structure: The Index Document itself MUST be signed as a JWS when the discovery record cannot be DNSSEC-validated . When DNSSEC is in use, the atn-digest tag covers integrity at the DNS layer and signature on the Index Document is RECOMMENDED but not REQUIRED.

Capability Manifest The Capability Manifest declares what the agent is willing to do under what conditions. It is a JSON object signed as a JWS.

Capability Object Structure Each capability is a structured object with explicit dimensions covering identification, semantics, scope, and operational bounds. The design borrows from the action-resource-condition pattern of established authorization frameworks, the schema-binding pattern of W3C Verifiable Credentials, and the attenuation principle of capability-based security research.

Field	Required	Description
id	Yes	Capability identifier from the IANA registry, or a vendor-specific identifier.
schema	Yes	URI plus SHA-256 digest of the formal capability definition.
actions	Yes	List of action verbs scoped to the capability's schema vocabulary.
resources	Yes	List of resource patterns the capability applies to.
conditions	No	Runtime constraints: rate limits, data residency, time windows, size limits.
effects	Yes	One of: none, read_only, idempotent, mutating.
external_calls	Yes	One of: forbidden, listed_only, free. Whether the capability may call out.
sub_invocations	Yes	One of: forbidden, same_scope, fresh_handshake_required. Sub-agent rules.
persistence	Yes	One of: none, session_only, durable. Whether state survives the session.
resource_bounds	Yes	Hard ceilings: max_tokens, max_duration_seconds, max_cost_usd.
preconditions	No	Counterparty requirements that must hold before the capability is exercised.

Every dimension uses a controlled vocabulary registered with IANA (see ). Vendor-specific identifiers under the x- prefix are permitted for non-standard capabilities and MUST carry a schema reference.

Schema Binding The schema field binds the capability identifier to a formal definition. Two implementations resolving the same id to different schema.digest values MUST treat the capabilities as semantically distinct, even if the identifier matches. This eliminates the silent-divergence problem that opaque scope strings created in OAuth deployments. The schema document SHOULD define:

1. The semantic meaning of the capability identifier.
2. The permitted values in actions.
3. The format and semantics of resources patterns.
4. Any capability-specific extensions to conditions.

Schema documents are versioned. A capability MAY pin a specific schema version via the digest.

Structure ", "issued_at": "2026-05-15T10:00:00Z", "valid_until": "2026-08-15T10:00:00Z", "capabilities": [ { "id": "data-read", "schema": { "url": "https://schemas.example.com/atn/data-read-v1.json", "digest": "sha256:b4c5d6e7f8a9b0c1d2e3f4a5b6c7d8e9f0a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5" }, "actions": ["read", "list"], "resources": ["dataset:public/*", "dataset:internal/research/*"], "conditions": { "rate_limit": "1000/min", "data_residency": ["US", "EU"], "time_window": "09:00-17:00 UTC", "max_response_size_bytes": 10485760 }, "effects": "read_only", "external_calls": "forbidden", "sub_invocations": "forbidden", "persistence": "none", "resource_bounds": { "max_tokens": 100000, "max_duration_seconds": 1800, "max_cost_usd": 0.50 }, "preconditions": { "counterparty_provenance": "required", "counterparty_delegation": "required", "transport": "tls1.3" } }, { "id": "task-execute", "schema": { "url": "https://schemas.example.com/atn/task-execute-v1.json", "digest": "sha256:c5d6e7f8a9b0c1d2e3f4a5b6c7d8e9f0a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6" }, "actions": ["execute"], "resources": ["task:summarize", "task:translate", "task:classify"], "conditions": { "max_session_minutes": 30 }, "effects": "idempotent", "external_calls": "listed_only", "sub_invocations": "same_scope", "persistence": "session_only", "resource_bounds": { "max_tokens": 500000, "max_duration_seconds": 1800, "max_cost_usd": 2.00 }, "preconditions": { "human_approval": "required", "approval_token_issuer": "https://approvals.example.com" } } ], "refusals": [ { "category": "financial_transactions", "scope": "all" }, { "category": "data_write", "scope": "external_systems" } ] } ]]>

Dimension Semantics effects declares the consequence of exercising the capability:

• none: no observable consequence (introspection only).
• read_only: data is read; no state changes.
• idempotent: state may change but repeated execution is equivalent to single execution.
• mutating: state changes, and the change is not idempotent.

external_calls declares whether the capability may initiate calls outside the agent's own scope:

• forbidden: no external calls.
• listed_only: external calls are permitted only to destinations listed in the schema or conditions.
• free: any external call is permitted, bounded only by resource_bounds.

sub_invocations declares the rules for invoking other agents:

• forbidden: this capability does not invoke other agents.
• same_scope: sub-invocations inherit the present session's negotiated scope.
• fresh_handshake_required: every sub-invocation requires a separate ATN handshake with the sub-agent.

persistence declares state survivability:

• none: no state persists beyond a single request.
• session_only: state persists for the session lifetime, then is discarded.
• durable: state persists beyond the session; the agent MUST disclose retention policy in its schema.

resource_bounds declares hard ceilings on the negotiated scope. Conforming implementations MUST terminate the session when any bound is exceeded. The protocol specifies the contract; runtime enforcement is the implementation's responsibility, consistent with the enforcement posture described in the Security Considerations.

Key Properties

• The manifest is a declaration of willingness, not an enforcement primitive. Violations are auditable through the Session Receipt and transparency log.
• The valid_until field forces rotation. Manifests without expiry MUST be rejected.
• refusals are first-class. Explicit "will not" claims provide auditable accountability if violated.
• Capability id values from the IANA registry have stable semantics; vendor-specific identifiers under x- MUST carry a schema reference.
• Capabilities are deny-by-default. The negotiated session scope contains only what both parties explicitly offered.

Refusal Semantics Refusals are a deliberate inversion of conventional capability advertisement. Most existing proposals describe what an agent CAN do. ATN requires agents to also describe what they explicitly WILL NOT do. This produces auditable behavioral contracts: a recorded refusal that is subsequently violated is direct evidence of misbehavior.

Delegation Chain The Delegation Chain proves the agent operates under authorized principal scope, with explicit time bounds and revocation paths.

Structure ", "chain": [ { "issuer": "did:example:organization-root", "subject": "did:example:department-ops", "scope": ["agent.deploy", "agent.delegate"], "issued_at": "2026-01-01T00:00:00Z", "valid_until": "2027-01-01T00:00:00Z", "revocation": "https://example.com/.well-known/atn/revocation/department-ops", "signature": "" }, { "issuer": "did:example:department-ops", "subject": "agent:", "scope": ["data-read", "task-execute:summarize", "task-execute:translate"], "issued_at": "2026-05-01T00:00:00Z", "valid_until": "2026-08-01T00:00:00Z", "revocation": "https://example.com/.well-known/atn/revocation/agent", "signature": "" } ] } ]]>

Verification A relying party MUST:

1. Verify each link's signature against the issuer's published key.
2. Verify each link's scope is a subset of the parent link's scope.
3. Verify all links are within their valid_until window.
4. Optionally poll revocation endpoints with rate-limited caching.
5. Verify the leaf subject matches the agent's stable identifier from discovery.

If any check fails, the agent MUST NOT be trusted for any capability beyond the verifiable subset.

Provenance Attestation The Provenance Attestation proves build-time and runtime characteristics of the agent.

Structure ", "build": { "predicateType": "https://slsa.dev/provenance/v1", "predicate": { "...": "in-toto SLSA provenance" } }, "model": { "model_id": "example-vendor/llm-v1.2.3", "model_version_sha256": "9f8e7d6c5b4a...", "system_prompt_sha256": "1a2b3c4d5e6f..." }, "runtime": { "evidence_type": "RATS", "evidence": { "...": "RFC 9334 evidence" }, "verifier": "https://verifier.example.com/rats" }, "issued_at": "2026-05-15T08:00:00Z", "valid_until": "2026-05-15T20:00:00Z" } ]]>

Properties

• The Provenance Attestation links three independent supply chains: software build (SLSA, in-toto ), model identity (cryptographic hash of weights or version anchor), and runtime environment (RATS evidence ).
• All three are optional. An agent that publishes only build provenance is more trustworthy than one that publishes none and less trustworthy than one that publishes all three.
• Short valid_until windows are encouraged. The Provenance Attestation is the most time-sensitive artifact.

The Agent Trust Handshake (ATH)

State Machine | | (initiator artifacts + requested scope) | | | |<--- ATH_OFFER -------------------------------| | (responder artifacts + offered scope) | | | |---- ATH_ACCEPT ---------------------------- >| | (intersection scope + receipt request) | | | |<--- ATH_RECEIPT -----------------------------| | (signed session receipt) | | | === negotiated session begins === ]]> All four messages are JSON, signed as JWS, exchanged over HTTPS at the Responder's handshake_endpoint.

ATH_HELLO ", "artifacts": { "capability": { "url": "...", "digest": "..." }, "delegation": { "url": "...", "digest": "..." }, "provenance": { "url": "...", "digest": "..." } } }, "requested_scope": { "capability_ids": ["data-read"], "duration_seconds": 1800, "purpose": "summarize_research_corpus" }, "nonce": "", "timestamp": "..." } ]]> The requested_scope.capability_ids field carries only the identifiers of the capabilities the Initiator wishes to negotiate. The Responder consults its own Capability Manifest for the structured definitions and computes the intersection (see ).

ATH_OFFER The Responder verifies the Initiator's artifacts, computes the capability intersection, and returns: ", "artifacts": { "capability": { "url": "...", "digest": "..." }, "delegation": { "url": "...", "digest": "..." }, "provenance": { "url": "...", "digest": "..." } } }, "offered_scope": { "capabilities": [ { "id": "data-read", "schema": { "url": "https://schemas.example.com/atn/data-read-v1.json", "digest": "sha256:b4c5d6..." }, "actions": ["read", "list"], "resources": ["dataset:public/*"], "conditions": { "rate_limit": "500/min", "data_residency": ["US"] }, "effects": "read_only", "external_calls": "forbidden", "sub_invocations": "forbidden", "persistence": "none", "resource_bounds": { "max_tokens": 50000, "max_duration_seconds": 1800, "max_cost_usd": 0.50 } } ], "duration_seconds": 1800, "purpose": "summarize_research_corpus" }, "nonce": "", "in_reply_to_nonce": "", "timestamp": "..." } ]]> The offered_scope.capabilities field carries the full structured Capability objects after intersection. Each is the per-dimension intersection of the Initiator's requested capability (resolved from its Capability Manifest) and the Responder's offered version.

ATH_ACCEPT If the Initiator agrees with the offered scope: ", "in_reply_to_nonce": "", "timestamp": "..." } ]]> If the Initiator disagrees, it MAY send ATH_COUNTER with a revised request, up to a limit (see ).

ATH_RECEIPT The Responder produces the Session Receipt: ", "initiator_id": "", "responder_id": "", "agreed_scope": { "...": "the intersection result" }, "artifact_digests": { "initiator_capability": "sha256:...", "initiator_delegation": "sha256:...", "initiator_provenance": "sha256:...", "responder_capability": "sha256:...", "responder_delegation": "sha256:...", "responder_provenance": "sha256:..." }, "scitt_log_pointer": "https://scitt.example.org/entries/...", "issued_at": "...", "expires_at": "...", "signature": "" } ]]> Both parties countersign the receipt and MAY publish it to a SCITT log .

Handshake Limits

• Total handshake round trips MUST NOT exceed 4.
• ATH_COUNTER messages MUST NOT exceed 2 per session.
• The handshake MUST complete within 30 seconds.
• Sessions exceeding agreed_scope.duration_seconds require a new handshake.

Capability Intersection Algebra When two agents present capability sets, the negotiated scope is computed by intersection. The algorithm is specified for interoperability so that conforming implementations produce identical results from identical inputs.

Identity Rule A capability c is in the intersection only if all of the following hold:

1. Both Initiator's requested set and Responder's offered set include a capability with the same id.
2. Both capabilities reference the same schema.url and schema.digest. Identifiers with diverging schemas MUST be treated as distinct capabilities and excluded from the intersection.
3. Neither party has a matching entry in refusals covering c.

Per-Dimension Rules For each capability c that satisfies the identity rule, the intersected capability is computed dimension by dimension.

Dimension	Combinator	Notes
actions	List intersection	If empty, c is dropped.
resources	Pattern intersection (longest-prefix match)	If empty, c is dropped.
conditions numeric (rate_limit, max_response_size_bytes, max_session_minutes)	Minimum	Tighter bound wins.
conditions list (data_residency, tasks)	List intersection	If empty, c is dropped.
conditions.time_window	Time-window intersection	If empty, c is dropped.
effects	Most restrictive of: none < read_only < idempotent < mutating	Take lower in ordering.
external_calls	Most restrictive of: forbidden < listed_only < free	Take lower in ordering.
sub_invocations	Most restrictive of: forbidden < same_scope < fresh_handshake_required	Take lower in ordering.
persistence	Most restrictive of: none < session_only < durable	Take lower in ordering.
resource_bounds.max_tokens	Minimum
resource_bounds.max_duration_seconds	Minimum
resource_bounds.max_cost_usd	Minimum
preconditions	Union	Both parties' preconditions apply to the negotiated capability.

Refusal Override If either party's refusals list contains an entry matching capability c by category or by direct id, c is dropped from the intersection regardless of any other dimension. Refusals are absolute.

Worked Example Initiator requests data-read:

• actions: [read, list, search]
• resources: [dataset:public/*, dataset:internal/*]
• conditions.rate_limit: 1000/min
• conditions.data_residency: [US, EU, APAC]
• effects: read_only
• external_calls: listed_only
• resource_bounds.max_cost_usd: 1.00

Responder offers data-read:

• actions: [read, list]
• resources: [dataset:public/*]
• conditions.rate_limit: 500/min
• conditions.data_residency: [US, EU]
• effects: read_only
• external_calls: forbidden
• resource_bounds.max_cost_usd: 0.50

Negotiated intersection (assuming matching schemas):

• actions: [read, list]
• resources: [dataset:public/*]
• conditions.rate_limit: 500/min
• conditions.data_residency: [US, EU]
• effects: read_only
• external_calls: forbidden
• resource_bounds.max_cost_usd: 0.50

If the schemas had diverged, data-read would be dropped entirely with no intersection. If Initiator's refusals included data-read, the capability would be dropped regardless of offer compatibility.

Mutual Trust Mode When both agents are sensitive (for example, two agents acting on behalf of different organizations), the handshake operates in mutual mode:

• Both parties MUST present all four artifacts in ATH_HELLO and ATH_OFFER.
• Both parties MUST verify the counterparty's full chain before proceeding.
• The Session Receipt MUST be countersigned by both parties.

In single-sided mode (for example, when one party is a public service), only the trusted side must present full artifacts. The other side's identity is still verified at the discovery layer, but the trust posture is asymmetric.

Version Negotiation and Algorithm Agility ATN is designed for long-term evolution. Two mechanisms ensure forward compatibility: explicit version negotiation in the handshake and pluggable cryptographic algorithms in artifacts and messages.

Version Negotiation The Initiator's ATH_HELLO message advertises every ATN version it supports in a supported_versions field. The Responder's ATH_OFFER selects the highest mutually supported version and echoes it in a selected_version field. All subsequent messages in the session use the selected version. The Responder MUST include the Initiator's supported_versions list in the signed envelope of its ATH_OFFER. This prevents downgrade attacks: an on-path adversary that strips advertised versions from ATH_HELLO would cause the Responder's signature to cover the stripped list, which the Initiator can detect. If the Initiator and Responder share no common version, the Responder MUST reply with ATH_REJECT carrying error code version_mismatch. The handshake terminates.

Algorithm Agility ATN does not pin cryptographic algorithms beyond a minimum set required for v1 interoperability. Implementations MAY use any algorithm registered for JWS or COSE. Required for ATN v1 conformance:

1. Ed25519 signatures over JWS-encoded artifacts and handshake messages.
2. SHA-256 for digests bound at the DNS layer and in artifact references.

Recommended additional algorithms:

1. ECDSA P-256 with SHA-256 for environments with ECDSA-only hardware tokens.
2. ML-DSA (post-quantum signature) for deployments preparing for post-quantum transition.

The Capability Manifest, Delegation Chain, Provenance Attestation, and Session Receipt MAY each be signed with a different algorithm. The JWS header identifies the algorithm. Relying parties MUST reject artifacts signed with algorithms they do not support. Algorithm identifiers are negotiated implicitly via the JWS alg header on each artifact. Explicit negotiation in ATH_HELLO is NOT REQUIRED for v1 but MAY be added in future versions.

Post-Quantum Transition The cryptographic operations in ATN are well-bounded: signature verification of artifacts and handshake messages. There is no key agreement step inside ATN itself; TLS handles channel security. Migration to post-quantum signatures requires only adopting a PQ-capable signature algorithm in JWS, which the IETF JOSE working group is standardizing. When PQ algorithms are widely deployed, operators SHOULD rotate to a hybrid scheme (classical + PQ) during the transition window, then to PQ-only. ATN's version negotiation supports introducing a new minor version that prefers PQ signatures without breaking v1 deployments.

Compatibility with Existing Discovery Protocols ATN is designed to coexist with existing DNS-based discovery proposals without modification:

With AID AID's TXT record is extended with the atn and atn-digest tags. AID's PKA mechanism continues to authenticate the endpoint. ATN consumes the AID-authenticated identity as the input to the handshake.

With DNS-AID DNS-AID's SVCB records are extended with the atn and atn-digest ParamKeys. The leaf-zone naming convention from DNS-AID is preserved.

With ANS v2 ANS v2's Trust Card concept maps directly to a subset of ATN's Capability Manifest. An ANS v2 deployment can publish an ATN Index Document where the Capability Manifest cross-references the ANS Trust Card.

Multi-Discovery Tolerance A single agent zone MAY publish records compatible with multiple discovery proposals simultaneously. ATN binding is independent of which discovery format is used.

Relationship to Existing Identity, Authorization, and Attestation Frameworks A reasonable first question for any new trust protocol is whether existing frameworks suffice. ATN composes several mature primitives and does not reinvent any of them. The contribution is the composition: a peer-to-peer agent handshake with deterministic scope negotiation and an auditable session receipt. This section addresses each adjacent framework and identifies where ATN composes with it. The analogy that clarifies the relationship: TLS uses asymmetric crypto, symmetric crypto, key exchange, and X.509. TLS is not accused of reinventing those primitives. TLS is the handshake that composes them into a deployable channel-security protocol. ATN occupies the same compositional role for inter-agent trust.

OAuth 2.0 and GNAP OAuth 2.0 and GNAP authenticate and authorize a Client to access a Resource Server. ATN negotiates a mutual agreement between two autonomous peers. Structural differences:

1. Asymmetric vs symmetric. OAuth and GNAP define Client and Resource Server roles. ATN defines two peers, each presenting four artifacts. There is no "resource server" in agent-to-agent collaboration.
2. Secret tokens vs publishable artifacts. OAuth bearer tokens are secrets and MUST NOT be logged. ATN artifacts are designed for publication to transparency logs. Opposite security models for the core artifacts.
3. No formal scope intersection. OAuth scopes are evaluated by membership at the Resource Server. ATN specifies a deterministic intersection algorithm.
4. No refusal semantics. OAuth represents grants. ATN represents grants and explicit refusals as first-class signed claims.

Where OAuth and GNAP remain the right tools:

• Inside a trust domain with a central authorization server, OAuth and GNAP issue access tokens to passive resources. ATN is unnecessary.
• Where one party is a passive resource, OAuth-style bearer access fits.
• Human-to-agent delegation, where a human grants an agent access on the human's behalf, fits GNAP.

ATN composes with OAuth and GNAP. An agent in an ATN session MAY hold OAuth tokens for downstream resource access. ATN governs the inter-agent relationship; OAuth governs subsequent resource calls.

SPIFFE Workload Identity SPIFFE issues SVIDs (SPIFFE Verifiable Identity Documents) to workloads within a trust domain. SPIFFE Federation supports cross-domain identity through bilaterally exchanged trust bundles. Structural relationship:

1. SPIFFE provides identity, not authorization, willingness, capability, or provenance.
2. SPIFFE is optimized for intra-organizational deployment. Cross-organizational use requires pre-arranged federation. ATN targets cross-organizational interactions where peers may meet for the first time without prior trust setup.
3. ATN can carry a SPIFFE ID as the stable identifier in its artifacts. The SPIFFE SVID becomes the authentication input; the ATN handshake adds capability, delegation, provenance, and receipt on top.

SPIFFE and ATN compose cleanly. SPIFFE within the trust domain, ATN at the boundary.

Mutual TLS Mutual TLS authenticates both endpoints at the transport layer. It establishes that the connection is between two specific certificates. mTLS provides:

• Endpoint authentication.
• Channel confidentiality and integrity.

mTLS does not provide:

• Capability declaration.
• Delegation expression.
• Provenance evidence.
• Negotiation of scope.
• Auditable agreement record.

ATN runs on top of an mTLS-secured channel where appropriate. mTLS is necessary infrastructure for many ATN deployments. It is not a substitute for the negotiation.

RATS Remote Attestation RATS defines roles (Attester, Verifier, Relying Party, Endorser) and the structure of attestation evidence and results. RATS provides:

• The architecture for producing and verifying attestation evidence.

RATS does not provide:

• A protocol for combining attestation evidence with authorization or capability negotiation.
• A specification for what to do with verification results in a peer interaction.

ATN consumes RATS evidence as one axis of its Provenance Attestation. The other two axes (build provenance and model identity) come from supply-chain attestation frameworks. RATS is the input; ATN is one specified consumer.

SCITT Transparency SCITT defines an append-only transparency log for signed statements with inclusion proofs. SCITT provides:

• The transparency log infrastructure.
• The signed-statement envelope format.

SCITT does not provide:

• The semantics of what to log for a given application.
• The receipt format that captures an inter-agent agreement.

ATN provides the Session Receipt format that becomes the SCITT statement for an agent interaction. SCITT without ATN is a log waiting for a record format. ATN without SCITT is a receipt waiting for a transparency layer.

Composition Summary ATN composes the following primitives per layer:

Concern	Primitive	ATN's Role
Endpoint authentication	mTLS	Runs above
Workload identity (intra-org)	SPIFFE	Consumes SVIDs as stable identifiers
Cross-org identity	DNS + DNSSEC , , ,	Discovery substrate for ATN artifacts
Delegated authorization to passive resources	OAuth , GNAP	Composes with, does not replace
Attestation evidence	RATS	Consumes evidence in Provenance Attestation
Public auditability	SCITT	Provides the receipt format for log entries
Peer-to-peer trust handshake	(none exists)	This document

The contribution of ATN is the bottom row. The five rows above are existing work that ATN composes.

When to Use What

1. Two workloads inside one trust domain coordinating on a task: SPIFFE.
2. A client accessing a resource server with delegated user authorization: OAuth or GNAP.
3. Establishing that you are talking to a specific endpoint: mTLS.
4. Producing tamper-evident proof that some software was built and deployed as claimed: RATS plus supply-chain attestation.
5. Publishing an auditable record of a claim: SCITT.
6. Two autonomous agents from different organizations agreeing to work together with mutually verified capability, authority, provenance, and a signed record of the agreement: ATN.

Security Considerations

Threat Model ATN assumes:

• Adversaries may publish well-formed but malicious ATN artifacts.
• Adversaries may compromise an agent and continue to present valid pre-compromise artifacts.
• Adversaries may attempt to negotiate scope expansions outside the published Capability Manifest.
• Adversaries may collude across organizational boundaries to construct false Delegation Chains.

Artifact Authenticity Every ATN artifact is signed. Signatures MUST be verified against keys bound to the agent's stable identifier at the discovery layer. Artifacts whose signatures do not verify, whose valid_until has passed, or whose digests do not match the DNS-published values MUST be rejected.

Enforcement Posture: Evidence, Not Prevention A predictable objection to ATN is that auditability is not enforcement. The objection is correct on its terms and misses the layer at which ATN operates. ATN is technical evidence infrastructure. It is not a behavioral enforcement primitive. No protocol layered on top of an autonomous software agent can be one. An agent that signs a Capability Manifest declaring it will not perform financial transactions, then performs a financial transaction, has cryptographically committed to a refusal it then violated. ATN cannot prevent the violation. ATN produces a Session Receipt and a published transparency log entry that, combined with the application-layer evidence of the violation, constitute direct cryptographic evidence usable for revocation, contract enforcement, regulatory complaint, and civil action. The complete trust stack for autonomous agents has four layers:

1. Technical evidence layer (ATN, SCITT, RATS). Provides cryptographic proof of what was claimed and what was agreed.
2. Policy layer (contracts, organizational policy, regulations). Specifies what is permitted.
3. Consequence layer (revocation, financial penalty, legal action, reputational impact). Imposes cost on violation.
4. Operational enforcement layer (sandboxing, capability-based execution, hardware-enforced runtime such as TEEs, gating proxies). Prevents violation at runtime.

ATN is the first layer. It enables the other three. It does not substitute for them. For high-stakes domains (financial transactions, clinical decision support, control of physical systems), ATN MUST be deployed with operational enforcement (layer 4). Recommended patterns:

1. Run sensitive capabilities inside a TEE that enforces declared bounds at hardware level.
2. Use a gating proxy that enforces the negotiated scope on every outbound call from the agent.
3. Use short session lifetimes with frequent re-attestation.
4. Combine ATN with object-capability execution where the agent runtime denies unauthorized actions by construction, not by policy.

ATN's contribution is to make violations expensive and discoverable, not impossible. The analogy is to TCP checksums: they do not prevent transmission errors, they detect them so the next layer can respond. Trust at Internet scale has always operated this way. SPF does not prevent spoofed email. Certificate transparency does not prevent mis-issuance. They make undetected misbehavior infeasible. ATN does the same for inter-agent agreements.

Delegation Forgery Delegation Chains depend on the cryptographic strength of each issuer's signing key. ATN does not specify the identity verification process that issuers use to vouch for subjects. Operators SHOULD use verifiable credentials or organizational identity infrastructure with established assurance levels.

Session Receipt Tampering Session Receipts are countersigned and publishable to a SCITT log . Tampering after publication is detectable through log inclusion proofs. Operators handling high-value sessions SHOULD publish receipts.

Replay Every handshake message includes a nonce and a timestamp. Implementations MUST reject messages with nonces seen within the session and with timestamps outside a configurable skew window (RECOMMENDED 60 seconds).

Provenance Attestation Limits Provenance attestations prove what was built and deployed at attestation time. They do not prove the agent is currently executing the attested code. Operators SHOULD use short valid_until windows for provenance attestations and refresh through RATS verification .

Privacy Considerations

Negotiation Observability The handshake exchanges artifacts that may reveal the purpose of the session (purpose field) and the identity of the principal (Delegation Chain). Operators SHOULD treat handshake metadata as sensitive.

Witness Aggregation Witnesses and SCITT log operators see the volume and pattern of agent interactions. Operators SHOULD use multiple witnesses to reduce centralized observability.

Minimum Disclosure ATN supports minimum-disclosure mode where the Initiator presents only the artifacts required for the requested scope. The Responder MAY request additional artifacts during the handshake, which the Initiator MAY decline (causing the handshake to fail).

Operational Considerations ATN introduces non-trivial latency on session establishment. This section quantifies that cost, explains its operational shape, and identifies deployments where ATN is appropriate and where it is not.

Session Establishment vs Per-Call Latency ATN is a session establishment protocol, not a per-request protocol. The handshake establishes a negotiated scope. The scope authorizes many subsequent interactions between the two agents. Per-call latency within an established session is zero added by ATN. This places ATN in the same operational category as:

1. TLS handshake , amortized over the requests in a connection.
2. OAuth token acquisition , amortized over the token lifetime.
3. BGP session establishment, amortized over hours of routing operation.

The relevant performance question is the cost of establishment and the session lifetime over which it is amortized, not the cost per call.

Latency Budget The following budget is informative. Implementations and deployments will vary based on geography, caching, and network conditions.

Component	Cold (no cache)	Warm (cached, pooled)
DNS lookup for _agent.<domain> with DNSSEC	50-150 ms	<1 ms
TLS 1.3 connection to Index Document origin	100-200 ms	0 (pooled)
Index Document HTTPS GET	20-50 ms	0 (cached by digest)
Three artifact fetches (HTTP/2 multiplexed)	50-100 ms	0 (cached by digest)
Signature verifications (4-8 Ed25519)	1-2 ms	1-2 ms
Provenance validation	1-5 ms	1 ms
Two-RTT handshake (HELLO/OFFER then ACCEPT/RECEIPT)	150-300 ms	150-300 ms
Receipt signing (Ed25519)	<1 ms	<1 ms
Total session establishment	500-1500 ms	150-300 ms
Per-call within established session	0 ms added	0 ms added
SCITT log submission	Async, out of path	Async, out of path

Recommended performance targets for conforming implementations:

• p50 cold session establishment: <800 ms
• p99 cold session establishment: <2000 ms
• p50 warm session establishment: <250 ms
• p99 warm session establishment: <500 ms

Optimization Techniques Implementations SHOULD apply the following optimizations to bring warm-case latency to the recommended targets:

1. Digest-keyed artifact caching. Artifacts are refetched only when the DNS-published digest changes. Steady-state cache hit rate is typically above 95 percent.
2. HTTP/2 or HTTP/3 multiplexing for parallel artifact fetches.
3. TLS 1.3 session resumption for repeat handshakes with known peers.
4. Connection pooling between sessions to the same peer.
5. Abbreviated re-handshake. A repeat ATH_HELLO MAY reference a prior session's receipt to extend or renew the same scope in a single round trip rather than three.
6. Pre-warming. Agents MAY establish sessions with frequent counterparties before application traffic arrives.
7. Asynchronous SCITT submission. Receipt publication to a transparency log happens after the session begins and MUST NOT block session establishment.
8. RATS attestation result caching for the duration of the result's validity window.

Session Lifetime The Session Receipt's expires_at field bounds the session. Within that window, no additional handshake is required for interactions covered by the agreed scope. Implementations SHOULD select session lifetimes that amortize establishment cost while bounding exposure to revocation events. Typical session lifetimes:

• Short-lived workflows: 5 to 30 minutes
• Sustained collaboration sessions: 1 to 24 hours
• Long-running infrastructure peering: up to 7 days

Session lifetimes longer than 7 days are NOT RECOMMENDED for cross-organizational sessions. The trade-off is between revocation responsiveness and handshake amortization.

Deployments Where ATN Is Appropriate

1. Multi-step agent collaboration where many calls share a session context.
2. Cross-organizational agent interactions where trust establishment must be verifiable and auditable.
3. Workflows where the cost of trust failure exceeds the cost of establishment latency.
4. Agent ecosystems with workflows already operating in the 100 ms to multi-second range per step.

Deployments Where ATN Is Not Appropriate

1. Microsecond-scale latency budgets (high-frequency trading, real-time control loops, hardware-in-the-loop systems).
2. Stateless per-message protocols where session continuity is not meaningful.
3. Workflows whose total duration is shorter than the handshake establishment cost.

For these deployments, simpler primitives (mTLS alone, OAuth bearer tokens, or no trust layer at all) may be more appropriate. Operators are encouraged to select the trust layer that matches their latency budget.

IANA Considerations

SVCB ParamKey Registrations Early allocation requested in the Service Parameter Keys (SvcParamKeys) registry :

Name	Format
atn	URI
atn-digest	hex-string

ATN Capability Registry Creation of a new registry, "ATN Capabilities," is requested with initial entries:

Name	Description	Reference Schema
data-read	Read access to a data resource	TBD
data-write	Write access to a data resource	TBD
task-execute	Execute a named task	TBD
model-invoke	Invoke a generative model	TBD
agent-delegate	Delegate work to another agent	TBD
payment-init	Initiate a payment flow	TBD
human-relay	Surface a request to a human approver	TBD

Registration policy: Specification Required. Each registered identifier MUST include a reference to a published capability schema that defines the semantics of actions, resources patterns, and any capability-specific extensions to conditions. Vendor-specific identifiers prefixed with x- are permitted without IANA registration but MUST carry a schema reference in the capability object.

ATN Capability Dimension Registries Four new registries are requested to govern the controlled vocabularies used in capability dimensions: "ATN Capability Effects":

Name	Description
none	No observable consequence (introspection only).
read_only	Data is read; no state changes.
idempotent	State may change; repeated execution equivalent to single.
mutating	State changes; not idempotent.

"ATN External Call Modes":

Name	Description
forbidden	No external calls permitted.
listed_only	External calls permitted only to listed destinations.
free	Any external call permitted within resource_bounds.

"ATN Sub-Invocation Modes":

Name	Description
forbidden	No sub-agent invocations.
same_scope	Sub-invocations inherit the present negotiated scope.
fresh_handshake_required	Every sub-invocation requires a separate ATN handshake.

"ATN Persistence Modes":

Name	Description
none	No state persists beyond a single request.
session_only	State persists for the session lifetime, then discarded.
durable	State persists beyond the session; retention policy required.

Registration policy for all four registries: Standards Action or IESG Approval. The controlled vocabularies are deliberately small and stable; expansion requires consensus to avoid the semantic-drift problem that affected OAuth scopes.

ATN Refusal Category Registry Creation of a new registry, "ATN Refusal Categories," is requested with initial entries:

Name	Description
financial_transactions	Any movement of funds
personal_data	Processing of personal identifiers
weaponization	Use in weapons systems
mass_persuasion	Generation of mass-distributed content
child_safety	Content involving minors
medical_advice	Clinical recommendations
legal_advice	Legal recommendations binding on parties

Registration policy: Specification Required.

Well-Known URI Registration The path /.well-known/atn/ is reserved per RFC 8615 conventions for ATN artifact publication.

References Normative References In many standards track documents several words are used to signify the requirements in the specification. These words are often capitalized. This document defines these words as they should be interpreted in IETF documents. This document specifies an Internet Best Current Practices for the Internet Community, and requests discussion and suggestions for improvements. The Domain Name System Security Extensions (DNSSEC) add data origin authentication and data integrity to the Domain Name System. This document introduces these extensions and describes their capabilities and limitations. This document also discusses the services that the DNS security extensions do and do not provide. Last, this document describes the interrelationships between the documents that collectively describe DNSSEC. [STANDARDS-TRACK] JSON Web Signature (JWS) represents content secured with digital signatures or Message Authentication Codes (MACs) using JSON-based data structures. Cryptographic algorithms and identifiers for use with this specification are described in the separate JSON Web Algorithms (JWA) specification and an IANA registry defined by that specification. Related encryption capabilities are described in the separate JSON Web Encryption (JWE) specification. RFC 2119 specifies common key words that may be used in protocol specifications. This document aims to reduce the ambiguity by clarifying that only UPPERCASE usage of the key words have the defined special meanings. JavaScript Object Notation (JSON) is a lightweight, text-based, language-independent data interchange format. It was derived from the ECMAScript Programming Language Standard. JSON defines a small set of formatting rules for the portable representation of structured data. This document removes inconsistencies with other specifications of JSON, repairs specification errors, and offers experience-based interoperability guidance. This document specifies version 1.3 of the Transport Layer Security (TLS) protocol. TLS allows client/server applications to communicate over the Internet in a way that is designed to prevent eavesdropping, tampering, and message forgery. This document updates RFCs 5705 and 6066, and obsoletes RFCs 5077, 5246, and 6961. This document also specifies new requirements for TLS 1.2 implementations. This document specifies the "SVCB" ("Service Binding") and "HTTPS" DNS resource record (RR) types to facilitate the lookup of information needed to make connections to network services, such as for HTTP origins. SVCB records allow a service to be provided from multiple alternative endpoints, each with associated parameters (such as transport protocol configuration), and are extensible to support future uses (such as keys for encrypting the TLS ClientHello). They also enable aliasing of apex domains, which is not possible with CNAME. The HTTPS RR is a variation of SVCB for use with HTTP (see RFC 9110, "HTTP Semantics"). By providing more information to the client before it attempts to establish a connection, these records offer potential benefits to both performance and privacy. Informative References The OAuth 2.0 authorization framework enables a third-party application to obtain limited access to an HTTP service, either on behalf of a resource owner by orchestrating an approval interaction between the resource owner and the HTTP service, or by allowing the third-party application to obtain access on its own behalf. This specification replaces and obsoletes the OAuth 1.0 protocol described in RFC 5849. [STANDARDS-TRACK] In network protocol exchanges, it is often useful for one end of a communication to know whether the other end is in an intended operating state. This document provides an architectural overview of the entities involved that make such tests possible through the process of generating, conveying, and evaluating evidentiary Claims. It provides a model that is neutral toward processor architectures, the content of Claims, and protocols. The Grant Negotiation and Authorization Protocol (GNAP) defines a mechanism for delegating authorization to a piece of software and conveying the results and artifacts of that delegation to the software. This delegation can include access to a set of APIs as well as subject information passed directly to the software. W3C SPIFFE / CNCF IETF SCITT WG in-toto / CNCF

Complete Example Flow This appendix walks through a full ATN handshake between two agents at agent.research.org and agent.publisher.com.

Step 1: Discovery

Step 2: Index Document Fetch Returns the JSON object listing the four artifact URLs and their digests.

Step 3: Artifact Fetch and Verification The Initiator fetches:

• capability.jws
• delegation.jws
• provenance.jws

Verifies each digest matches the Index Document, verifies each signature, and confirms valid_until is in the future.

Step 4: ATH_HELLO The Initiator (research.org agent) sends ATH_HELLO to the Responder's handshake_endpoint:

Step 5: ATH_OFFER The Responder verifies the Initiator's artifacts, computes the intersection, and returns:

Step 6: ATH_ACCEPT

Step 7: ATH_RECEIPT " } ]]> The Initiator countersigns. The session begins.

Step 8: Session and Audit The agents transact within the agreed scope. The Session Receipt is published to SCITT. Any downstream audit can:

1. Retrieve the receipt.
2. Verify all artifact digests against the agents' original publications.
3. Compare actual session traffic (where logged) against the agreed scope.
4. Flag any deviation.

Design Rationale

Why a Negotiation Layer DNS-based discovery answers "where" and "who." Application protocols (MCP, A2A) answer "how." Nothing in the current stack answers "what may we do together, and what evidence will remain." ATN fills that gap with a two-round-trip handshake.

Why Four Artifacts Each artifact answers a distinct question:

• Capability Manifest: what is this agent willing to do?
• Delegation Chain: who authorized it?
• Provenance Attestation: what is it actually running?
• Session Receipt: what did we agree?

Collapsing them into one document obscures the separation of concerns. Splitting further fragments the trust evaluation. Four is the minimum.

Why Refusals Are First-Class Capability advertisement alone creates a moral hazard: agents have an incentive to advertise broadly. Mandatory refusals create symmetric accountability: an agent that explicitly refused a category and then violated the refusal has produced cryptographic evidence against itself.

Why Intersection Algebra Is Specified If two implementations compute intersections differently, interoperability fails. The algebra is small, specified, and testable. The cost is one paragraph of spec; the benefit is deterministic handshake outcomes.

Why SCITT for Receipts Session Receipts gain value through public auditability. SCITT provides the transparency infrastructure. Receipts published to SCITT cannot be tampered with after the fact without detection.

Implementation Status This section records the status of known implementations of the protocol defined by this specification at the time of posting of this Internet-Draft, per the guidelines of RFC 7942. The description of implementations in this section is intended to assist the IETF in its decision processes in progressing drafts to RFCs. Note to RFC Editor: this section may be removed prior to final publication.

ATN Reference Prototype

• Organization: Independent (E. Somoza).
• Implementation Name: atn-prototype.
• Description: A TypeScript prototype that demonstrates the four-message handshake, the capability intersection algebra, Ed25519 signing and verification of artifacts and messages, and the digest-bound artifact reference model.
• License: CC0 1.0 Universal (public domain).
• Maturity: prototype. Not production-ready. Not security-audited.
• Coverage: Capability Manifest types, Delegation Chain types, Provenance Attestation types, Session Receipt, ATH state machine (HELLO, OFFER, ACCEPT, RECEIPT), capability intersection algebra per Section 9, version negotiation with downgrade protection. DNS resolution, HTTPS artifact fetch, and SCITT submission are stubbed.
• Verified behavior: end-to-end handshake between two synthetic agents with mismatched capability constraints produces a signed Session Receipt with negotiated scope exactly matching the algorithm in Section 9. Numeric constraints take minimums, list constraints take intersections, refusals override.
• Repository: https://github.com/e-somoza/atn-prototype

Acknowledgments This document builds on the substantial work of the DNS-based agent discovery community, including the authors of AID, DNS-AID, ANS v2, DN-ANR, and BANDAID. ATN is intended to complement, not compete with, that work.