AGTP-API: Verbs, Paths, Endpoints, and Synthesis

1.1. Why the Contract Belongs in the Protocol

The architectural goal AGTP is designed around is runtime contract negotiation between autonomous agents (RCNS). An agent encountering an AGTP server must be able to:¶

• Discover what endpoints the server exposes, with full structural detail, in a single call.¶

• Reason about whether an endpoint matches its intent based on the endpoint's declared semantic block.¶

• Verify that its identity carries the authority required to invoke the endpoint.¶

• Propose new endpoints when its needs are not met by existing ones.¶

• Compose endpoints across servers while preserving authority through the composition.¶

None of this works if the contract lives in middleware. A protocol that does not know what an interaction means cannot govern it. When an agent invokes a method on a path, the protocol must verify identity, confirm permissions cover the method-on-path combination, confirm server policy allows the combination, validate the request body against the endpoint's input schema, apply scope checks, log the invocation, and -- if the request triggers synthesis -- ensure each composed step preserves authority. Each of these checks requires the protocol to know what the method-on-path combination means.¶

This is the load-bearing claim of AGTP-API. The contract is a first-class protocol concept rather than something inferred from documentation, OpenAPI specs, or framework conventions. Operational concerns -- rate limiting, observability, caching, custom authentication strategies, request transformation, deployment topology -- remain middleware territory because they are operational decisions specific to a deployment. The protocol governs contracts; middleware governs operations. Neither layer pretends to be the other.¶

1.2. Document Lineage

AGTP-API supersedes two previously-separate drafts:¶

• AGIS (the Agentive Grammar and Interface Specification, draft-hood-independent-agis-01) -- deprecated. Retained as a historical citation source for the empirical research informing the verb-class semantics. ("Verb" here is the linguistic category — the catalog is a vocabulary of action-intent verbs. See Section 2.)¶

• AGTP-Methods (the proposed Standard Extended Method Vocabulary, draft-hood-agtp-standard-methods-01) -- folded into this document. There is no separate method-catalog specification.¶

The previously-proposed AMG (Agent Method Grammar) draft is also deprecated; it never reached publication. The grammar framing is replaced by the contract framing throughout. This document assumes the reader understands AGTP [AGTP] v07 or later.¶

1.3. Requirements Language

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD 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.¶

3.1. Catalog Structure

The catalog is a versioned JSON document with five top-level keys:¶

version

Semantic version of the catalog as a whole. See Section 4.¶

embedded

The 12 protocol-floor verbs every AGTP server MUST support: six cognitive (QUERY, DISCOVER, DESCRIBE, SUMMARIZE, PLAN, PROPOSE) and six mechanics (EXECUTE, DELEGATE, ESCALATE, CONFIRM, SUSPEND, NOTIFY).¶

legacy

The five HTTP-style verbs (GET, POST, PUT, DELETE, PATCH), each carrying a preferred mapping to its canonical AGTP replacement (FETCH, CREATE, REPLACE, REMOVE, MODIFY). Legacy verbs are opt-in per server through the manifest's policies.methods block. See Section 3.6.¶

categories

The taxonomy used to classify verbs: discovery, retrieval, analysis, transaction, modification, creation, notification, mechanics, domain_spanning.¶

verbs

The catalog of approved verbs, each with categories, description, and optional deprecation metadata (deprecated_in, removed_in, successor).¶

The catalog is published in machine-readable form at the URL above and is intentionally maintained as an open living artifact rather than a closed IANA registry. Curation is performed by the AGTP-API maintainers in public; the version of the catalog in effect at any given moment is declared in the catalog_version field of every server manifest. A closed IANA registry would freeze the verb vocabulary at the cadence of IETF process; AGTP-API verbs need to evolve at the cadence of agent deployments. The IANA registries that AGTP-API does request (Section 17) cover the structural status codes, the media types, and the response headers — all of which benefit from IANA's conservative cadence. The catalog itself does not.¶

3.2. Method Rules

Method names MUST be uppercase ASCII single tokens matching the regex ^[A-Z]+$, length 3 to 32 characters inclusive. Method names MUST appear in the catalog, or in the legacy set when the server's policies.methods.legacy opts into legacy verbs.¶

The catalog is curated to contain only imperative-form action-intent verbs that:¶

• Express a command or action, not a state, an attribute, or a noun (RESERVE is valid; RESERVATION is not).¶

• Carry a single intent (verbs with compound intent are decomposed into separate verbs).¶

• Are not HTTP method names.¶

These are curatorial principles applied during verb addition, not runtime checks performed by AGTP servers. Servers validate verbs against the catalog membership and the lexical regex; semantic class adherence is the responsibility of the catalog maintainers.¶

3.3. Custom Methods

Servers MAY accept additional methods beyond the catalog by listing them in the manifest's policies.methods block. Custom methods follow the same lexical rules in Section 3.2 but are server-specific and are not interoperable across organizations without explicit acceptance. Custom methods are not added to the catalog; they exist only within the publishing server's surface.¶

3.4. Catalog Evolution Overview

The catalog evolves under semver discipline:¶

• Patch (1.0.0 → 1.0.1) — description or category metadata changes only.¶

• Minor (1.0.0 → 1.1.0) — verbs added; existing verbs unchanged. Minor revisions MAY introduce new deprecation flags on existing verbs.¶

• Major (1.0.0 → 2.0.0) — verbs removed or renamed.¶

Per-verb deprecation uses three optional fields: deprecated_in (catalog version that flagged the verb), removed_in (version scheduled for removal), and successor (recommended replacement). A deprecated verb remains admitted by the dispatcher; the response carries an AGTP-Catalog-Warning advisory header. See Section 4 for the full lifecycle, runtime semantics, removal behavior, cross-server version negotiation, and the agtp-catalog-diff tooling.¶

3.5. Catalog Version Exposure

The catalog version is exposed on the server manifest under catalog_version and catalog_versions_supported so clients can detect mismatches between their local catalog and the server's. See Section 8.¶

3.6. Legacy Verb Support

AGTP-API recognizes GET, POST, PUT, DELETE, and PATCH as legacy verbs with reframed semantics for transitional deployments. Legacy verb support is opt-in per server through the manifest's policies.methods block:¶

legacy_verbs: enabled

When legacy verbs are enabled, the server MUST treat them as follows:¶

• GET maps to QUERY for read operations against a known path.¶

• POST maps to EXECUTE for state-changing operations.¶

• PUT maps to EXECUTE with idempotency_key.¶

• DELETE maps to EXECUTE with action: "delete".¶

• PATCH maps to EXECUTE with action: "modify".¶

These mappings are conveniences for migration. Legacy verbs MUST NOT be used in new endpoint definitions; servers SHOULD deprecate legacy support over time.¶

4.1. Versioning Scheme

The method catalog is versioned as semver MAJOR.MINOR.PATCH:¶

• PATCH — editorial changes (clarifications, typo fixes, description rewordings). No verb additions, removals, or deprecations.¶

• MINOR — additions and deprecations. Verbs may be added; verbs may be marked deprecated. No verbs are removed in a minor revision.¶

• MAJOR — removals. Verbs marked deprecated in earlier minor revisions may be removed.¶

Servers and agents using a given catalog MAJOR version are guaranteed that no verb available in that MAJOR will disappear without warning; removal requires a MAJOR bump.¶

4.2. Deprecation Metadata

Each verb in the catalog MAY carry deprecation metadata:¶

{
  "name": "AUDIT_LEGACY",
  "category": "compute",
  "deprecated_in": "1.2.0",
  "removed_in": "2.0.0",
  "successor": "AUDIT"
}

deprecated_in:

The catalog version at which the verb became deprecated.¶

removed_in:

The anticipated catalog version at which the verb will be removed. MAY be absent if removal is anticipated but the target version has not been chosen.¶

successor:

A catalog verb name that callers should migrate to. MAY be absent if no single successor applies; clients should consult the verb's description for migration guidance.¶

4.3. Manifest Catalog Declaration

Every server manifest MUST declare:¶

catalog_version:

The catalog version the server validates incoming method names against.¶

catalog_versions_supported:

A list of catalog versions the server can validate against. The current catalog_version MUST appear in this list. Multi- version support is anticipated; v00 servers commonly list a single version.¶

Agents reading a manifest learn which catalog version they should use for outbound requests against this server. Mismatches are handled at request time per the runtime semantics below.¶

4.4. Runtime Deprecation Signaling

When a server processes a request whose method is deprecated in the catalog version the server is using, the server MUST include the AGTP-Catalog-Warning header on the response:¶

AGTP-Catalog-Warning: deprecated; successor=AUDIT; removed_in=2.0.0

Header fields:¶

• deprecated — keyword token indicating the warning class.¶

• successor=NAME — the recommended replacement verb. Omitted if the catalog metadata does not declare a successor.¶

• removed_in=VERSION — the target removal version. Omitted if not declared.¶

The header is advisory, not a rejection. The request MUST still process. Agents and CLIs that surface the header to their users allow operators to schedule migration before forced removal.¶

A server that receives a request whose method is in the catalog at all (deprecated or not) MUST NOT return 459 Method Violation. 459 is reserved for verbs that are not in the catalog at all. Deprecated verbs are still in the catalog.¶

4.5. Removal Behavior

When a verb is removed in a MAJOR catalog revision, servers upgrading to that revision MUST:¶

1. Reject inbound requests for the removed verb with 459 Method Violation, citing the catalog version the server is now using.¶

2. Refuse to load endpoint TOMLs declaring removed verbs at startup, logging which endpoints are affected.¶

3. Invalidate any active synthesis plans whose steps reference removed verbs, returning a structured failure on the next synthesis tick.¶

4. Skip policies.methods entries (allow, disallow, redirects) that reference removed methods, logging each skip.¶

This is graceful degradation: the server starts, serves the endpoints that are still valid, and surfaces what was lost. The operator decides whether to roll back, fix the affected endpoints, or accept the reduced surface.¶

4.6. Cross-Server Version Negotiation

When an agent talks to a server whose catalog_version differs from the agent's own preferred version, the agent SHOULD:¶

1. Read catalog_versions_supported from the manifest.¶

2. If the agent's preferred version is in the list, use it; the server will validate against that version.¶

3. If not, downgrade to the highest mutually-supported version.¶

4. If no overlap exists, abandon the connection or escalate to the principal.¶

Servers MAY support multiple catalog versions concurrently; servers MAY support only the single version they declare in catalog_version. Agents adapt to the server's capabilities.¶

4.7. Tooling

A reference agtp-catalog-diff tool is published alongside the method catalog. The tool diffs two catalog versions, reports added / removed / newly-deprecated methods, and MAY scan a deployment directory (endpoint TOMLs, recipe registry, agtp-server.toml) to identify which deployment artifacts would break under a candidate catalog upgrade. See Appendix A.¶

Tool invocation:¶

agtp-catalog-diff old.json new.json [--against-deployment <directory>]

The tool produces a structured diff showing methods added in the new catalog, methods removed, methods newly deprecated, path-grammar conflicts (existing endpoint paths containing newly-added method names), endpoint conflicts (existing endpoints declaring removed methods), recipe conflicts (existing recipes referencing removed methods), and policy conflicts (existing policy directives referencing removed methods).¶

Operators SHOULD run this tool before upgrading a server's catalog version. The tool's exit code indicates whether breaking changes exist: 0 (no breakage), 1 (breakage in deployment context), 2 (parse errors).¶

The tool is implementation-provided; alternative implementations MAY ship equivalents. The diff format is specified by the implementation's documentation rather than normatively by this section.¶

4.8. Multi-Version Catalog Support

Future revisions of AGTP-API may specify how servers can validate against multiple catalog versions simultaneously, allowing graceful migration during transitions. The catalog_versions_supported manifest field is reserved for this purpose.¶

For v00 conforming servers, catalog_versions_supported MUST be a single-element array containing only the server's current catalog_version.¶

4.9. Catalog Publishing

The canonical AGTP method catalog is published at https://agtp.io/api/methods.json. The catalog is versioned independently of AGTP-API itself; a server may implement AGTP-API v00 against catalog version 1.0.0 today, and against catalog version 1.1.0 after a minor catalog update.¶

Servers MAY validate against alternative catalogs (industry- specific catalogs, internal-only catalogs, etc.) by configuring their build to use a different catalog document. The catalog_version exposed in the manifest reflects whichever catalog the server actually validates against.¶

Catalog updates follow a coordinated release process: the publisher announces the upcoming version, operators run the diff tool against their deployments, the new catalog is published, operators upgrade their servers. Specific governance for the canonical catalog (proposal process, approval cadence, etc.) is documented separately and is out of scope for this specification.¶

5.1. Path Rules

A request path MUST begin with /. A request path MUST NOT end with / unless the path is exactly / (the bare root). No path segment MAY match an approved AGTP method name (case-insensitive, after stripping - and _). Parameter segments wrapped in {...} are exempt from the method-token check.¶

Any violation MUST return status 460 Endpoint Violation with the offending segment in the response body.¶

5.2. Path Grammar ABNF

The path component of the request line follows this grammar:¶

path           = "/" [ segment-nz *( "/" segment ) ]
segment        = *segment-char
segment-nz     = 1*segment-char
segment-char   = unreserved / pct-encoded / sub-delims
               / ":" / "@"
parameter      = "{" parameter-name "}"
parameter-name = 1*( ALPHA / DIGIT / "_" )
unreserved     = ALPHA / DIGIT / "-" / "." / "_" / "~"
pct-encoded    = "%" HEXDIG HEXDIG
sub-delims     = "!" / "$" / "&" / "'" / "(" / ")"
               / "*" / "+" / "," / ";" / "="
                                        ; per RFC 3986

A segment MAY be a literal segment (matching segment-nz with no { or } characters) or a parameter segment (matching the parameter production exactly). Mixing literal and parameter forms within a single segment (e.g., prefix-{id}) is NOT supported in v00; future revisions MAY specify mixed-form templates.¶

Parameter names within a path MUST be unique. A path containing two parameters with the same name MUST be rejected at endpoint registration time.¶

5.3. Permissiveness

Mixed case, underscores, hyphens, and any segment depth are valid. The protocol does not impose style preferences. Casing conventions, kebab-vs-snake-case, and parameter-naming style are operator decisions.¶

Path templates support typed parameters in curly braces:¶

/customers/{customer_id}
/orders/{order_id}/line-items/{line_item_id}
/products/category/{category_slug}

The {name} form is the only template syntax v00 admits. URI Template ([RFC6570]) syntax (e.g., {?query,page}, {+reserved}, {#fragment}) is NOT supported; servers encountering a path template that uses anything other than the basic {name} form MUST reject the endpoint at registration time. The conservative single-form rule keeps endpoint dispatch deterministic and predictable; future revisions MAY broaden the template grammar.¶

Template parameters MUST be declared in the endpoint's input_schema as JSON Schema properties of the same name. A template parameter that is not declared in input_schema MUST cause registration to fail. Parameter values arrive at the handler as part of the validated input alongside body parameters.¶

5.4. Path Pattern Matching

Servers match an incoming request path against registered endpoints in three passes:¶

1. Exact match. If the request path exactly equals a registered literal path (no template parameters), the matching endpoint is selected.¶

2. Template match. If no exact match, the request path is matched against registered templates. A template matches when the path has the same number of segments as the template and every literal segment matches exactly while every {param} segment captures the corresponding request segment as a parameter value.¶

3. No match. If neither pass matches, the server MUST return 404 Not Found.¶

When multiple templates would match the same request path, the server MUST select the template with the fewest parameter segments (most specific wins). If two templates have the same number of parameter segments and both match, registration MUST have failed at startup with a path-ambiguity error; the dispatcher MUST NOT be reached in this state.¶

5.5. Query Strings

Paths MAY carry a query string after a ?:¶

AGTP/1.0 SCHEDULE /meeting?date=050526&attendees=alice%2Cbob

The wire layer splits the request-target at the first ? per [RFC3986]: the path is everything from the leading / up to the first ? or end-of-line; the query string is everything after the first ?. Path-grammar enforcement (method-name leakage check, template matching) applies to the path component only; query strings carry no semantic constraints from this grammar.¶

Query parameters merge into the request input alongside body parameters before schema validation. On key conflicts, body parameters take precedence. Repeated query keys collapse to the last value; multi-valued shapes ride in the body.¶

Servers MUST percent-decode query parameter values per [RFC3986] Section 2.1 before schema validation.¶

5.6. Fragments

URI fragments (#anchor) are REJECTED at the wire layer. AGTP traffic carries no client-side anchoring use case; a # in the request line is always malformed. Servers MUST return 400 Bad Request with error code invalid-request-line for any request line containing #.¶

5.7. No Reserved Path Prefixes

AGTP reserves no path prefixes. There is no /.well-known/ discovery surface and no /_agtp/ namespace. AGTP is its own protocol; HTTP-style discovery conventions do not apply.¶

5.8. Server Metadata via AGTP Methods

Server-internal metadata is exposed via AGTP methods at AGTP-native paths, registered as built-in endpoints alongside operator-authored endpoints. Servers MUST expose DISCOVER /methods, returning a compact inventory of every endpoint the server has registered:¶

[
  {
    "method": "QUERY",
    "path": "/catalog",
    "description": "Returns the current product catalog."
  },
  {
    "method": "BOOK",
    "path": "/room",
    "description": "Books a room for the named guest at the named property."
  },
  {
    "method": "DISCOVER",
    "path": "/methods",
    "description": "Lists all registered endpoints on this server."
  }
]

The response is a JSON array of objects, each with method, path, and description. The verb is DISCOVER because the response carries endpoints to talk to (the AGTP semantics for DISCOVER, see [AGTP]) rather than data to consume; agents use the response to drive follow-on invocations against the listed endpoints.¶

The DISCOVER /methods response is a lightweight inventory; agents that need full endpoint definitions (semantic blocks, schemas, deprecation metadata, handler types) MUST retrieve the full server manifest via target-less DISCOVER (Section 8). The built-in exists as a cheap, low-bandwidth probe for endpoint shape — useful for clients that need to know what to ask for without loading the full manifest.¶

Built-in endpoints appear in the server manifest like any other endpoint (including DISCOVER /methods itself, which is one of the entries returned by its own invocation), are subject to the same dispatcher gates (catalog, path grammar, policy), and MAY be overridden by operator-authored TOML at the same (method, path) pair. Operators that override a built-in are responsible for preserving the documented response contract.¶

Future revisions of this specification MAY define additional built-in endpoints (e.g., QUERY /catalog-version for catalog version negotiation). Built-ins are protocol-level conventions, not deployment-level extensions.¶

6.1. Endpoint Definition

An endpoint is the structural unit an AGTP server exposes. Every endpoint is identified by the pair (method, path) and carries the fields specified below.¶

{
  "method": "BOOK",
  "path": "/room",
  "description": "Books a room for the named guest at the named property.",
  "namespace": "reservations",

  "semantic": {
    "intent": "Reserve a room for the named guest at the named property.",
    "actor": "agent",
    "outcome": "A confirmed reservation_id is returned for the guest.",
    "capability": "transaction",
    "confidence": 0.85,
    "impact": "irreversible",
    "is_idempotent": false
  },

  "input_schema": {
    "$schema": "https://json-schema.org/draft/2020-12/schema",
    "type": "object",
    "properties": {
      "guest_id": { "type": "string", "format": "uuid" },
      "room_id":  { "type": "string" },
      "arrival":  { "type": "string", "format": "date" },
      "departure":{ "type": "string", "format": "date" }
    },
    "required": ["guest_id", "room_id", "arrival", "departure"],
    "additionalProperties": false
  },

  "output_schema": {
    "$schema": "https://json-schema.org/draft/2020-12/schema",
    "type": "object",
    "properties": {
      "reservation_id": { "type": "string", "format": "uuid" }
    },
    "required": ["reservation_id"],
    "additionalProperties": true
  },

  "errors": ["room_unavailable", "invalid_dates"],

  "handler": { "type": "registered_function" },

  "required_scopes": ["booking:room", "calendar:write"],

  "deprecated": {
    "deprecated_in": "2.1.0",
    "removed_in": "3.0.0",
    "successor": { "method": "RESERVE", "path": "/rooms" }
  }
}

6.2. Required Fields

Every endpoint MUST carry: method, path, description, semantic, input_schema, output_schema, errors, handler. The remaining fields (namespace, required_scopes, deprecated) are optional.¶

6.3. Field Semantics

method:

An AGTP method drawn from the approved catalog at the server's declared catalog_version, or a custom method declared in the server's policies.methods. See Section 3.3. The method name satisfies the lexical rules of Section 3.2.¶

path:

A path conforming to the path grammar of Section 5.¶

description:

Operator-facing prose. A single sentence describing what the endpoint does. Distinct from the semantic block: description is for humans reading the manifest; the semantic block is for agents reasoning about applicability.¶

namespace:

Optional free-form grouping label. Servers exposing many endpoints MAY organize them by namespace (e.g., "reservations", "billing", "inventory") for catalog navigation. Namespace has no protocol-level effect.¶

semantic:

The structural metadata block defined in Section 7. Required by every endpoint with all required fields populated.¶

input_schema:

A Draft 2020-12 JSON Schema [JSON-SCHEMA] describing the request body. MUST declare "type": "object" and MUST set "additionalProperties": false; see Section 13.3. Query-string parameters (Section 5.5) are merged into the body before validation; the same schema covers both. Schemas are inline in v00; cross-endpoint schema sharing is anticipated in a future revision.¶

output_schema:

A Draft 2020-12 JSON Schema describing the successful response body. SHOULD set "additionalProperties": true so handlers can return forward-compat fields without bumping the contract; see Section 13.3.¶

errors:

An array of named error condition strings (e.g., "room_unavailable", "invalid_dates"). The dispatcher returns these as 422 Unprocessable Entity responses with structured bodies identifying the condition. The error name is the protocol-level identifier; status-code mapping is a response-construction concern, not part of the contract. The list MAY be empty for endpoints with no business-level errors. Composition handlers MUST include "composition_failed"; external-service handlers MUST include the upstream-failure codes specified in Section 12.3.¶

handler:

An object declaring the binding kind: {"type": "registered_function" | "composition" | "external_service"}. The handler's full reference (Python dotted path, recipe name, upstream URL, etc.) is server-internal implementation detail and MUST NOT appear in the manifest. See Section 8.9 and Section 12.¶

required_scopes:

An array of scope identifiers the invoking agent MUST declare in its Authority-Scope. Empty or omitted means no scope check; the endpoint inherits the server's default authority requirement. Insufficient scope returns 455 Scope Violation per [AGTP].¶

deprecated:

Optional per-endpoint deprecation metadata. See Section 6.4.¶

6.4. Endpoint Deprecation

The optional deprecated block follows the same shape as catalog-level deprecation:¶

"deprecated": {
  "deprecated_in": "2.1.0",
  "removed_in": "3.0.0",
  "successor": { "method": "RESERVE", "path": "/rooms" }
}

Field semantics:¶

deprecated_in:

The server's agtp_api_version (or operator-chosen version string) at which the endpoint became deprecated.¶

removed_in:

Optional. The version at which the endpoint will be removed. MAY be absent if removal is anticipated but the target version has not been chosen.¶

successor:

Optional. An object with method and/or path indicating the replacement endpoint callers should migrate to. Either field MAY be absent: successor.method alone signals "use the same path with a different method"; successor.path alone signals "use the same method with a different path"; both together signal a full replacement endpoint.¶

An endpoint MAY be deprecated even when its method and path remain valid in the catalog and grammar. Operators deprecate endpoints when they migrate callers from one (method, path) pair to another, independent of catalog evolution.¶

The dispatcher MUST stamp an AGTP-Endpoint-Warning advisory header on every response from a deprecated endpoint:¶

AGTP-Endpoint-Warning: deprecated; successor=RESERVE /rooms; removed_in=3.0.0

The header rides alongside AGTP-Catalog-Warning (Section 4) when both apply; both are advisory and the request still processes. The successor token renders as METHOD /path when both fields are present; just one is emitted bare. removed_in and successor are omitted from the header when the endpoint's deprecated block does not declare them. The header fires on success and failure responses alike so callers always see the warning regardless of how the invocation resolves.¶

7.1. Required Semantic Block Fields

intent:

A natural-language statement of what the endpoint accomplishes. Single sentence. Imperative form. Subject is the actor; the verb reflects the endpoint's method; the object is the path resource.¶

actor:

A short identifier of the agent class that should invoke this endpoint. Freeform string; servers MUST NOT reject endpoints whose actor is not in a fixed enum. Suggested values for interoperable authoring include agent, human, system, customer, staff, admin; domain-specific values (e.g., merchant, auditor, patient) are equally valid. The field helps agents recognize whether an endpoint applies to their role but carries no protocol-level enforcement.¶

outcome:

A natural-language statement of the post-condition produced by successful invocation. Distinct from intent: intent is what the endpoint sets out to do; outcome is what is true after.¶

capability:

The high-level effect class of the endpoint. One of the catalog category values: discovery, retrieval, analysis, transaction, modification, creation, notification, mechanics, domain_spanning. The same taxonomy used to classify methods in the catalog (Section 3); an endpoint inherits its capability class from the category of the method it exposes, though operators MAY override when the endpoint's specific behavior fits a different category.¶

confidence:

A scalar on the closed interval 0.0 to 1.0 expressing how confident the server is that an agent invoking this endpoint with satisfying input will produce the declared outcome. Endpoints that are deterministic and well-tested score near 1.0; endpoints that involve nondeterministic resources or partial failures score lower. Agents MAY use the value to decide whether to ESCALATE for clarification before invoking. The scale is advisory and not normatively defined beyond the endpoint bounds.¶

impact:

The reversibility of the endpoint's effect. One of informational, reversible, irreversible. informational endpoints return data without changing server state. reversible endpoints change state in ways that can be undone via a compensating method (e.g., CANCEL, REFUND). irreversible endpoints commit changes that cannot be programmatically undone (e.g., sending an email, charging a payment, shipping goods). Governance frameworks MAY use this field to apply differential authorization, gating irreversible endpoints behind stricter scope checks.¶

is_idempotent:

Boolean. Whether repeated invocation with identical input produces identical effect. Affects retry safety: idempotent endpoints MAY be retried freely; non-idempotent endpoints require an Idempotency-Key header per [AGTP] to be safely retried.¶

8.1. Purpose

The server manifest is the machine-readable document an AGTP server publishes to declare its identity, version posture, endpoint catalog, hosted agents, hosted protocols, and operational policies. It is the authoritative answer to "what does this server offer."¶

8.2. Manifest Structure

{
  "agtp_version": "1.0",
  "agtp_api_version": "1.0",
  "document_version": "v2",
  "catalog_version": "1.0.0",
  "catalog_versions_supported": ["1.0.0"],

  "server": {
    "server_id": "agents.acme.com",
    "domain": null,
    "operator": "Acme Retail",
    "contact": "ops@acme.com",
    "supported_features": ["endpoint-registry", "synthesis"],
    "issued": "2026-01-15T09:00:00Z",
    "updated": "2026-04-15T09:00:00Z"
  },

  "embedded_methods": [ /* twelve floor methods */ ],
  "custom_methods": [ /* server-specific methods declared in policies.methods */ ],
  "endpoints": [ /* array of endpoint definitions, projected */ ],

  "agent_disclosure": "public",
  "hosted_agents": [ /* agent identity references */ ],
  "agent_disclosure_notice": null,

  "apis": [ /* legacy HTTP APIs this server fronts, if any */ ],
  "hosted_protocols": [ /* MCP, A2A, ACP, etc. */ ],

  "policies": {
    "wildcards_accepted": false,
    "anonymous_discovery": true,
    "scope_required_for_invocation": true,
    "synthesis_enabled": true,
    "max_synthesis_depth": 10
  },

  "manifest_signature": null
}

The manifest carries three distinct version fields at the top level:¶

agtp_version:

The AGTP wire-protocol version the server speaks (currently "1.0", corresponding to AGTP v07 of [AGTP]). Pinned to the base protocol, separate from contract-layer evolution.¶

agtp_api_version:

The AGTP-API contract-layer version the server implements (currently "1.0" for v00 of this document). Carries the endpoint primitive, semantic block, status code, and synthesis semantics the server adheres to.¶

document_version:

This specific manifest's revision string. Operator-controlled; bumped on every meaningful change to the manifest content. Permits agents to detect cache staleness without comparing the full body.¶

catalog_version:

The method catalog version the server validates incoming method names against. See Section 4.¶

catalog_versions_supported:

List of catalog versions the server can validate against. The current catalog_version MUST appear in this list. Multi-version support is anticipated in a future revision but is not normatively specified in v00.¶

8.3. Server Block

The server block carries operator-facing identity and timestamps:¶

server_id:

The canonical identifier for this server. MAY be a hostname, an agtp:// URI, or an operator-defined string. The value MUST be stable across manifest revisions.¶

domain:

Optional operational host the server is reachable at, if different from server_id. MAY be null for servers whose server_id is itself a resolvable host.¶

operator:

The human-readable name of the organization operating the server.¶

contact:

An operator email address for security and operational issues.¶

supported_features:

A non-normative array of feature tokens describing optional capabilities the server exposes (e.g., "endpoint-registry", "synthesis", "openapi-export"). Agents and operators MAY use these tokens for capability negotiation. AGTP-API does not define a canonical feature-token registry; tokens are operator-defined.¶

issued:

ISO 8601 timestamp of when the manifest was first issued. Preserved across manifest revisions; identifies the manifest's origin in time.¶

updated:

ISO 8601 timestamp of the most recent manifest update. Bumped on every regeneration.¶

8.4. Methods Inventory

The embedded_methods array carries the twelve protocol-floor methods every AGTP server supports (see [AGTP] §6). The custom_methods array carries any server-specific methods declared in policies.methods (Section 3.3). The custom_methods field MAY be omitted when empty.¶

These arrays are non-normatively populated; agents that want a complete picture of what the server invokes SHOULD consult the endpoints array, which is the authoritative source for what methods are bound on what paths.¶

8.5. Agents Disclosure

The three agent-related fields describe what agents the server hosts and how they are exposed:¶

agent_disclosure:

One of "public", "authenticated", "private". "public" means the hosted_agents array is populated and visible to all callers. "authenticated" means the array is populated only for callers presenting a verified identity. "private" means the array is empty in this manifest; the server does not disclose what agents it hosts.¶

hosted_agents:

Array of agent identity references the server hosts. Format is operator-defined; typical entries include a canonical Agent-ID and a brief description.¶

agent_disclosure_notice:

Optional human-readable note explaining the server's disclosure policy. Useful when agent_disclosure is "authenticated" or "private" and the operator wishes to clarify how a caller can request elevated disclosure.¶

8.6. Protocol and API Surface

apis:

Optional array of legacy HTTP API descriptors this server fronts (e.g., entries pointing at OpenAPI specifications). Servers that expose a parallel HTTP surface alongside AGTP list those HTTP endpoints here; the presence of entries in this array is the signal that the server operates a parallel HTTP face.¶

hosted_protocols:

Optional array of higher-level agent protocols the server hosts on top of AGTP (e.g., MCP, A2A, ACP). Each entry is an object with the protocol name and the AGTP method+path it is reachable via. See the composition section in [AGTP].¶

8.7. Policies Block

The policies block carries five normative operational toggles:¶

wildcards_accepted:

One of true, false. Whether the server accepts ad-hoc method invocations against undeclared endpoints. See Section 5.8.¶

anonymous_discovery:

One of true, false. Whether unauthenticated callers may retrieve the manifest and invoke read-only built-in endpoints. Servers serving traffic from open networks typically set this to true; servers in trusted-network deployments MAY set this to false to require authenticated identity even on discovery.¶

scope_required_for_invocation:

One of true, false. Whether every endpoint invocation requires the agent to carry an Authority-Scope header. true is the default; the implementation MUST NOT invoke a handler when scope is required and absent. Servers with no scope-gated endpoints MAY set this to false.¶

synthesis_enabled:

One of true, false. Whether the server participates in PROPOSE-and-synthesis runtime contract negotiation. When false, PROPOSE requests MUST be refused with 463 Proposal Rejected and the reason synthesis-disabled.¶

max_synthesis_depth:

A positive integer specifying the maximum number of composed steps a synthesized endpoint MAY combine. Plans exceeding this depth are refused; the server falls through to counter-proposal or 463. Default: 10.¶

8.8. Manifest Signature

The manifest body SHOULD carry a manifest_signature field holding a JWS envelope over the canonical JSON encoding of the manifest body (with the manifest_signature field itself excluded from the signing input). Clients MAY verify signatures when present and SHOULD warn when absent.¶

A future revision of this specification will require manifest_signature (MUST); reference implementations are expected to ship signing support in alignment with that revision. The signing algorithm, canonical encoding rules, and key-publication mechanism are specified separately in [AGTP-CERT].¶

For v00, the manifest_signature field MAY be null or absent; a manifest without a valid signature MUST NOT be rejected by clients on signature grounds alone.¶

8.9. Endpoint Projection

The endpoints array in the manifest is a projection of the server's internal endpoint catalog. The projection differs from the full endpoint definition in one important way: the handler's type-specific reference field (function, recipe, or url) and any other binding-specific implementation details are dropped, leaving only the binding kind under handler.type.¶

Internal endpoint definition (on disk, in TOML or equivalent):¶

{
  "method": "BOOK",
  "path": "/room",
  "description": "...",
  "semantic": { ... },
  "input_schema": { ... },
  "output_schema": { ... },
  "errors": [ ... ],
  "required_scopes": [ ... ],
  "handler": {
    "type": "registered_function",
    "function": "rooms.handlers.book_room"
  }
}

Manifest projection (over the wire):¶

{
  "method": "BOOK",
  "path": "/room",
  "description": "...",
  "semantic": { ... },
  "input_schema": { ... },
  "output_schema": { ... },
  "errors": [ ... ],
  "required_scopes": [ ... ],
  "handler": { "type": "registered_function" }
}

The type-specific reference field (function, recipe, or url) and any other binding-specific implementation fields are server-internal and MUST NOT appear in the manifest projection. Agents need to know the binding kind to reason about expected latency, retry behavior, and attribution; agents do not need to know which specific function, recipe, or upstream realizes the binding.¶

Servers MUST apply this projection on every manifest response.¶

8.10. Manifest Retrieval

Agents retrieve the server manifest using the AGTP DISCOVER method without an Agent-ID header. The absence of an Agent-ID header distinguishes a server-level discovery request (returns the manifest) from an agent-level discovery request (returns endpoint candidates matching specified criteria).¶

The response MUST carry Content-Type application/vnd.agtp.manifest+json. The manifest SHOULD be signed; see the Manifest Signature subsection above.¶

8.11. Manifest Caching

Servers SHOULD include Cache-Control and ETag headers on manifest responses. Agents MAY cache manifests according to those directives. The server.updated timestamp and the document_version field permit agents to detect cache staleness without revalidation: a changed document_version means the manifest content has changed in a way the operator considers significant.¶

9.1. Purpose

The method policy declares per-server constraints on which methods are accepted, which legacy verbs are opted in, and which methods on which paths are redirected to their canonical form. The policy is the small, declarative complement to the endpoint catalog: the catalog says what an endpoint does; the policy says what the server accepts in the first place.¶

In v00, the method policy is a sub-block of the manifest's policies (Section 8) carried at policies.methods. There is no separate policy file format; the policy is part of the server's configuration and is exposed on the manifest like any other operational policy.¶

9.2. Policy Shape

The policies.methods block is a JSON object with four optional fields:¶

allow:

Either the string "*" (the server admits every method in the catalog plus any method it has registered an endpoint for), or an array of method names (the server admits only the listed methods plus the embedded twelve floor methods). Default: "*".¶

disallow:

Array of method names the server explicitly rejects. Disallowed methods are rejected with 405 Method Not Allowed even when they would otherwise be permitted by allow. Useful for narrowing the surface without enumerating an exhaustive allow list. Default: empty array.¶

legacy:

Either an array of legacy HTTP verb names the server opts into (e.g., ["GET", "POST"]), the string "*" (all five legacy verbs accepted), or the string "NONE" (no legacy support). Granular per-verb opt-in supersedes the binary enabled/disabled toggle. Default: "NONE".¶

redirects:

Array of method-and-path redirect entries. Each entry has the shape:¶

{
  "from_method": "BOOK",
  "from_path": "/room",
  "to_method": "RESERVE",
  "to_path": "/room"
}

¶

Either from_path or to_path MAY be omitted, signaling "redirect on method-only without path matching" or "redirect to the same path as the request." When a request matches a redirect entry, the server processes it as if the to_method/to_path had been requested instead.¶

9.3. Example

A representative TOML deployment configuration:¶

[policies.methods]
allow = "*"
disallow = ["PATCH", "TRANSFER"]
legacy = ["GET"]

[[policies.methods.redirects]]
from_method = "BOOK"
from_path = "/room"
to_method = "RESERVE"
to_path = "/room"

The equivalent manifest projection:¶

"policies": {
  "methods": {
    "allow": "*",
    "disallow": ["PATCH", "TRANSFER"],
    "legacy": ["GET"],
    "redirects": [
      {
        "from_method": "BOOK",
        "from_path": "/room",
        "to_method": "RESERVE",
        "to_path": "/room"
      }
    ]
  }
}

9.4. Catalog-Graceful Skip Semantics

When the server loads policies.methods and an entry under allow, disallow, or redirects references a method removed from the catalog version the server has loaded, the entry MUST be skipped at load time with an operator-visible warning. The server continues to operate; the policy entry has no effect. This permits operators to upgrade the catalog without simultaneously rewriting the policy block to remove all references to retired methods.¶

The legacy field continues to enforce the strict set of five HTTP verbs (GET, POST, PUT, DELETE, PATCH); typos or non-HTTP verbs in legacy MUST be rejected as a hard configuration error.¶

9.5. Discovery

Agents discover the server's method policy by retrieving the server manifest via target-less DISCOVER (Section 8). The policies.methods block is part of the manifest body and is covered by the manifest signature (Section 8) when present.¶

Agents that need only an endpoint inventory (which methods on which paths are exposed, with one-line descriptions) MAY invoke DISCOVER /methods (Section 5.8) for a lightweight response that bypasses the full manifest weight.¶

10.1. Purpose

PROPOSE is the AGTP method that triggers runtime contract negotiation. An agent that needs an endpoint the server does not advertise submits a PROPOSE request carrying an AGTP-API-conformant endpoint specification; the server evaluates the proposal against the contract layer and either instantiates the endpoint as a session-scoped artifact or refuses.¶

PROPOSE is one of the twelve floor methods defined in [AGTP]. Its parameter table is specified there. This document specifies the synthesis semantics that govern server-side evaluation.¶

10.2. Negotiation Flow

Step 1:  Agent retrieves the server manifest (which carries the
         method policy as `policies.methods`)
Step 2:  Agent determines the needed endpoint is not declared
Step 3:  Agent constructs an AGTP-API-conformant endpoint definition
Step 4:  Agent sends PROPOSE with the endpoint definition in body
Step 5a: Server accepts and synthesizes
         Server evaluates contract conformance and capability envelope
         Server composes the endpoint from existing endpoints
         Server returns 263 Proposal Approved with the
           AGTP-API endpoint definition for the synthesized endpoint
         Synthesis-ID matches original proposal
         Synthesized endpoint is session-scoped by default
         Agent MAY invoke the endpoint immediately
Step 5b: Server rejects (463)
         Server returns 463 Proposal Rejected with structured reason
         Response SHOULD reference manifest entries that satisfy a
           similar capability if available
         Agent MAY modify proposal and retry (maximum 3 turns)
         After 3 rejections agent MUST ESCALATE

10.3. Server-Side Synthesis Evaluation

On receipt of a PROPOSE request, the server MUST evaluate:¶

1. Method conformance. Is the proposed method in the AGTP-API approved method catalog, or is it a custom method declared in the server's policies.methods (Section 3.3)? If not, return 459 Method Violation.¶

2. Path conformance. Does the proposed path satisfy Section 5? If not, return 460 Endpoint Violation.¶

3. Semantic block conformance. Does the proposed semantic block carry all required fields with valid values? If not, return 400 Bad Request.¶

4. Schema conformance. Are the input and output schemas valid JSON Schema documents? If not, return 400 Bad Request.¶

5. Authority sufficiency. Does the proposing agent's declared Authority-Scope cover the proposed authority requirements? If not, return 262 Authorization Required with body type scope-required.¶

6. Capability envelope. Can the server actually synthesize the proposed endpoint from its existing capabilities? If not, return 463 Proposal Rejected with reason: composition-impossible.¶

7. Policy acceptance. Does the server's policy permit the proposed method-and-path combination? If not, return 463 with reason: policy-refused.¶

If all evaluations pass, the server synthesizes the endpoint and returns 263 Proposal Approved.¶

10.4. Synthesis-Composed Endpoints

When a server synthesizes an endpoint, the server is composing the proposed behavior from existing endpoints internally. The synthesized endpoint is exposed to the agent as a single endpoint, but the server's handler implementation invokes one or more underlying endpoints to fulfill the request.¶

Authority MUST be preserved through the composition. The agent's Authority-Scope MUST be sufficient for every underlying endpoint the synthesis invokes. If a composed step would require authority the agent does not hold, the synthesis MUST fail at evaluation time (step 5) rather than at execution time.¶

10.5. Session-Scoped vs Persistent Synthesis

By default, a synthesized endpoint is session-scoped: it exists for the duration of the AGTP session that proposed it and is removed when the session ends. The agent MAY request persistent synthesis by setting persistent: true in the PROPOSE body; the server MAY honor or refuse the persistence request independently of the synthesis decision. Persistent synthesis subjects the proposed endpoint to the server's regular endpoint catalog and is reflected in subsequent manifest retrievals.¶

12.1. Composition

A composition handler exposes a derived endpoint whose behavior is realized by orchestrating one or more other endpoints declared on the same server. No application code is invoked; the server's synthesis runtime executes the composition steps in order, threading outputs into subsequent inputs while preserving authority through the composition.¶

Composition handlers reference a named recipe:¶

"handler": {
  "type": "composition",
  "recipe": "audit-via-query-and-summarize"
}

The recipe value is the name of a recipe defined in the server's recipe registry (typically in an agtp-recipes.toml file loaded at server startup). Recipes are reusable units of composition logic; one recipe MAY back multiple endpoints. Inline composition steps are not supported in v00; every composition endpoint MUST reference a registered recipe by name.¶

A composition handler whose recipe value does not resolve to a recipe registered at the server MUST cause registration to fail.¶

Composition endpoints MUST include composition_failed in their errors list. This is the named error returned when a composition step fails (any underlying step returns a non-success status). The error response carries details with the recipe identifier, the failed step number, the underlying step method, the underlying status, and any captured outputs from preceding steps.¶

Authority MUST be preserved through the composition. The agent's Authority-Scope MUST be sufficient for every underlying endpoint the composition invokes. Authority sufficiency is verified at registration time when possible (when step paths are statically known) and at invocation time for parameterized step paths.¶

This is the binding type most commonly produced by PROPOSE + synthesis: a synthesized endpoint is a composition of existing endpoints, and MAY be promoted to a permanent declared composition by adding its recipe to the server's recipe registry and writing an endpoint TOML that references it by name.¶

12.2. Registered Function

"handler": {
  "type": "registered_function",
  "function": "rooms.handlers.book_room"
}

The endpoint binds to a Python function loaded by the server. The function value is a dotted path identifying the module and the callable within it; the server imports the module at registration time and resolves the function attribute. Single dotted-path form is normative; separate module and function fields are not supported in v00.¶

The function receives the validated input and returns the response payload. The handler API contract is specified separately (see the reference implementation's handler documentation). This is the binding type used for endpoints whose logic exceeds what composition can express.¶

12.3. External Service

"handler": {
  "type": "external_service",
  "url": "https://internal.acme.tld/booking",
  "method": "POST",
  "headers": {
    "Authorization": "Bearer ${BOOKING_API_TOKEN}",
    "Content-Type": "application/json"
  },
  "input_transform": {
    "guest_id": "guestId",
    "room_type": "roomType"
  },
  "output_transform": {
    "confirmation_code": "confirmationNumber"
  },
  "error_map": {
    "404": "room_not_found",
    "409": "room_unavailable"
  },
  "timeout_seconds": 30
}

The endpoint binds to an HTTPS-reachable external service. The server proxies the request to the external endpoint and returns the response. Useful for AGTP frontends to existing HTTP-based services. The external service is not part of the AGTP contract; the AGTP server presents a contract-compliant facade.¶

External service fields:¶

url:

The fully-qualified URL of the upstream HTTP endpoint. The URL MUST use the https:// scheme. Plain HTTP MUST NOT be accepted at registration time.¶

method:

The HTTP method to invoke on the upstream. One of GET, POST, PUT, PATCH, DELETE, HEAD, OPTIONS. REQUIRED.¶

headers:

Optional map of HTTP headers to send to the upstream on every request. Values MAY include ${VAR} placeholders that are resolved against process environment variables at server startup. Unresolved placeholders MUST cause registration to fail.¶

input_transform:

Optional map renaming AGTP request fields to the upstream's expected field names. Maps AGTP-side name to upstream-side name. Fields not in the map pass through unchanged. v00 supports key-rename only; value transformation is anticipated in a future revision.¶

output_transform:

Optional map renaming upstream response fields to AGTP-side names. The map is inverted at response time: the upstream-side name is the source, the AGTP-side name is the target. Fields not in the map pass through unchanged.¶

error_map:

Map of upstream HTTP status code (as a string) to AGTP error code declared in the endpoint's errors list. Servers SHOULD configure error_map for every external_service binding; without it, upstream HTTP errors surface to agents as generic upstream_error. Every value in error_map MUST appear in the endpoint's errors; registration MUST fail otherwise.¶

timeout_seconds:

Positive number specifying the upstream request timeout. Servers SHOULD set this explicitly. Default if absent: 30 seconds. Production deployments SHOULD NOT rely on the default; explicit timeouts make upstream-dependency posture inspectable.¶

External service endpoints MUST declare the following five upstream-failure error codes in their errors list, in addition to any business-level errors:¶

Without these declarations, transport-level upstream failures would surface to the calling agent as opaque server errors. Mandating them makes the failure surface inspectable.¶

Agent identity is NOT propagated to the upstream by default. This is a security default: an upstream HTTP service has no mechanism to validate AGTP agent credentials, and forwarding them would expose principal identity to systems outside the AGTP trust boundary. Servers wishing to propagate identity to a specific upstream MAY do so via explicit header configuration in the headers map (e.g., a server-issued bearer token), but MUST NOT forward AGTP-Agent-ID or AGTP-Principal-ID by default.¶

13.1. Single Tier

A conforming AGTP server implements:¶

1. The full endpoint primitive (Section 6).¶

2. Method validation against the AGTP-API approved catalog at the server's declared catalog_version.¶

3. Path validation against the AGTP-API path grammar.¶

4. Method policy enforcement at dispatch time per the policies.methods sub-block of the manifest (Section 9).¶

5. A server manifest retrievable via DISCOVER without an Agent-ID header, carrying agtp_version, agtp_api_version, document_version, catalog_version, and catalog_versions_supported per Section 8.¶

6. The endpoint projection of Section 8.9: manifest responses expose handler.type only, not the full handler binding.¶

7. PROPOSE handling with the synthesis evaluation flow specified in this document, gated by policies.synthesis_enabled.¶

8. The structural rejection status codes 404, 405, 459, 460, plus the contract-negotiation codes 261, 262, 263, 463.¶

9. The semantic block on every published endpoint.¶

10. Schema validation for input and output, with the strictness asymmetry of Section 13.3 (inputs strict, outputs permissive by default).¶

11. The AGTP-Catalog-Warning response header on every response whose method is deprecated in the catalog version in use.¶

12. The AGTP-Endpoint-Warning response header on every response from a deprecated endpoint (Section 6.4).¶

13. The DISCOVER /methods built-in endpoint defined in Section 5.8, returning a compact inventory of the server's registered endpoints.¶

There is no halfway tier. Servers that implement only method dispatch without endpoints are not AGTP-conformant; they are something else, perhaps a method-based RPC layer that happens to use AGTP framing.¶

13.2. Validation Timing

Conforming servers perform two distinct categories of validation, which SHOULD occur at different times.¶

Manifest validation at startup. Before accepting traffic, a conforming server SHOULD validate every endpoint declared in its manifest:¶

• The method appears in the AGTP-API approved catalog (or is a custom method declared in the server's policies.methods).¶

• The path conforms to the path grammar of Section 5.¶

• The semantic block carries every required field.¶

• The input and output schemas are syntactically valid JSON Schema documents.¶

• The handler binding (Section 12) is resolvable: a composition's referenced steps exist; a registered_function's module and function are importable; an external_service's URL is well-formed and reachable (or marked deferred).¶

A server that detects manifest validation failures at startup SHOULD reject the offending endpoints with clear logs and SHOULD NOT silently ignore them. Servers MAY choose to fail startup entirely on manifest validation failure or to start serving with the valid subset of endpoints; the choice is a deployment policy decision.¶

Request validation at runtime. Once serving, the server validates each inbound request against the endpoint's input schema before invoking the handler:¶

• Input body validates against the endpoint's input_schema. On failure, the server returns 422 Unprocessable Entity with a schema validation error report.¶

• The agent's Authority-Scope covers the endpoint's required_scopes declarations. On failure, the server returns 455 Scope Violation per [AGTP].¶

• The verb-and-path combination is exposed by the server's current policy. On failure, the server returns the appropriate structural rejection code (404, 405, 459, or 460).¶

Validation MUST complete before handler invocation. Servers MUST NOT invoke the handler with input that has not satisfied schema validation; this guarantee is what allows handler authors to assume their context structure is well-formed.¶

The startup-vs-runtime split is a recommendation, not a strict requirement. Servers MAY defer startup validation if their endpoint catalog is dynamic (e.g., a PROPOSE-driven server that only synthesizes endpoints on demand). Such servers MUST still validate at the moment an endpoint is registered or synthesized, just not at process start.¶

13.3. Schema Strictness

Schema validation is asymmetric between inputs and outputs.¶

Input schemas MUST be validated with additionalProperties: false. Fields not declared in the schema MUST cause validation to fail with 422. This catches typos and version mismatches early and prevents agents from sending unrecognized fields under the mistaken assumption that the server will silently ignore them.¶

Output schemas SHOULD be validated with additionalProperties: true. A handler that returns more fields than the schema declared is forward-compatible: agents can ignore extra fields. A handler that returns missing required fields fails validation either way. Servers MAY opt into strict output validation as a deployment policy.¶

This asymmetry is normative because it has interoperability consequences. An agent submitting a request to two different servers can rely on both rejecting unrecognized input fields. An agent consuming a response from two different servers can rely on forward-compatible field handling without the version of one server breaking integration with the other.¶

13.4. Conformance Test Suite

The AGTP-API conformance test suite is maintained at:¶

https://agtp.io/api/conformance

The test suite exercises every conformance requirement in this section. Servers claiming AGTP-API conformance MUST pass the test suite at the version they declare in their manifest's agtp_api_version field.¶

14.1. OpenAPI as an Authoring Source

A developer with an existing OpenAPI specification can convert it to AGTP-API endpoint declarations. The HTTP method maps to an AGTP verb (with translation: GET maps to QUERY or DISCOVER depending on intent; POST maps to EXECUTE, BOOK, PURCHASE, or another action verb depending on the operation). The path remains. The request schema becomes input_schema. The response schema becomes output_schema. Errors become the errors list. The OpenAPI description becomes the foundation of the semantic block's intent.¶

Conversion tooling MAY automate this translation with reasonable defaults, flagging cases where verb mapping requires human judgment. The AGTP-API approved method catalog (Section 3) is the canonical source for HTTP-verb-to-AGTP-verb conversions: tooling MUST select target verbs from the list rather than inventing ad-hoc names, which is what makes converted endpoints interoperable with native AGTP-API endpoints.¶

Servers SHOULD NOT treat machine-converted endpoints as authoritative until a human reviewer has confirmed the semantic block fields, particularly intent, outcome, and impact. Conversion tooling has no reliable way to infer impact from HTTP semantics; that is judgment work.¶

14.2. OpenAPI as an Export Target

AGTP servers MAY render their endpoint catalog as an OpenAPI document for HTTP consumers that do not yet speak AGTP. The reverse translation: AGTP verb to HTTP method (mappings are imperfect; some AGTP verbs have no clean HTTP analog and are best exposed via the legacy verb compatibility mode of Section 3.6); path remains; schemas pass through unchanged.¶

The exported OpenAPI document is a compatibility surface, not the authoritative description of the service. Servers exporting OpenAPI SHOULD mark the export as derived from the AGTP-API manifest and SHOULD publish a link to the manifest in the OpenAPI info section.¶

14.3. OpenAPI as a Coexistence Layer

A server MAY run AGTP and HTTP side by side. The AGTP endpoint's handler is the same code that the HTTP endpoint invokes. OpenAPI documents the HTTP face; the AGTP-API manifest documents the agent face. The handler does not know which protocol invoked it. This deployment posture is signaled in the manifest by populating the apis array (Section 8) with descriptors pointing at the server's OpenAPI documents.¶

Coexistence is the path most large organizations will take during transition: AGTP is added as a parallel face without requiring modification to existing HTTP infrastructure. Implementation guidance for coexistence deployments is provided in Appendix A.¶

16.1. Verb-Path Tampering

A malicious server could declare endpoints whose semantic blocks misrepresent the actual behavior. Agents that rely on semantic blocks for authorization decisions MUST also enforce Authority-Scope checks against actual outcomes, not declared outcomes.¶

16.2. Synthesis Authority Preservation

The synthesis evaluation step (step 5 in Section 10.3) MUST verify that authority is preserved through composition. A synthesis that elevates authority by composing endpoints with sufficient individual scopes into a behavior the agent should not be permitted is a known threat. Servers MUST reject syntheses where the composed behavior exceeds what the agent's Authority-Scope permits, regardless of whether the individual steps are within scope.¶

16.3. Method Policy Tampering

The method policy is carried as the policies.methods sub-block of the server manifest and is covered by the manifest signature (Section 8) when present. Agents that retrieve the manifest over AGTP transport (which mandates TLS 1.3 or higher per [AGTP]) and verify the manifest signature when present are protected against policy substitution. Agents that accept manifest content without signature verification are vulnerable to substitution of the method policy along with the rest of the manifest.¶

16.4. Wildcard Abuse

The wildcards: true consent on the agent identity document and wildcards_accepted: true on the server policy are mutual consents that enable ad-hoc method invocation. Agents SHOULD NOT declare wildcards: true by default; servers SHOULD NOT declare wildcards_accepted: true for any operation with non-trivial impact tier. The 262 Authorization Required response with body type wildcards-required is the safe default.¶

16.5. Verb List Trust

Agents trust the AGTP-API approved method catalog as published at https://agtp.io/api/methods.json. The catalog MUST be served over HTTPS with content integrity verification. Agents MAY pin the catalog version they have validated and refuse upgrades without explicit operator approval.¶

17.1. AGTP-API Status Code Assignments

The following codes are registered in the AGTP Status Code Registry with AGTP-API as the authoritative reference:¶

• 261 Negotiation In Progress¶

• 262 Authorization Required¶

• 263 Proposal Approved¶

• 405 Method Not Allowed (AGTP-API semantics)¶

• 459 Method Violation¶

• 460 Endpoint Violation¶

• 463 Proposal Rejected¶

17.2. Media Type Registrations

This document defines two media types and requests their registration in the IANA Media Types registry. Full registration templates are provided in the appendices. The master registry of all AGTP-family media types is maintained in [AGTP].¶

application/vnd.agtp.manifest+json

The AGTP server manifest format. Carried on responses to a server-level DISCOVER request (DISCOVER without an Agent-ID header). The registration application will be filed concurrent with publication of this document.¶

application/vnd.agtp.endpoint+json

The AGTP-API endpoint definition format. Used in PROPOSE request bodies and in 263 Proposal Approved response bodies. The registration application will be filed concurrent with publication of this document.¶

Until the IANA registrations complete, the reference implementation MAY continue to emit application/json for manifest and endpoint responses. Implementations are encouraged to emit the AGTP-specific types as soon as they are accepted by IANA so that content negotiation and MIME-based routing can distinguish AGTP traffic from generic JSON.¶

17.3. AGTP-API Method Catalog Reference

This document does NOT request establishment of an IANA AGTP-API Verb List Registry. As specified in Section 3.1, the verb list is intentionally maintained as an open living artifact at https://agtp.io/api/methods.json rather than as a closed IANA registry, so that the catalog can evolve at the cadence of agent deployment rather than the cadence of IETF process. The structural status codes, the media types, and the response headers below are appropriate for IANA registration; the method catalog is not.¶

17.4. AGTP-API Response Headers

This document requests registration of the following response headers in the IANA HTTP Field Name Registry (or equivalent AGTP-specific registry where applicable):¶

AGTP-Catalog-Warning:

Advisory header surfaced on responses where the request method is deprecated in the catalog version the server is using. Format and semantics specified in Section 4.¶

AGTP-API-Version:

Header surfacing the AGTP-API specification version a manifest or endpoint definition conforms to.¶

20.1. Normative References

[AGTP]

Hood, C., "Agent Transfer Protocol (AGTP)", Work in Progress, Internet-Draft, draft-hood-independent-agtp-07, 2026, <https://datatracker.ietf.org/doc/html/draft-hood-independent-agtp-07>.

[JSON-SCHEMA]

Wright, A., Andrews, H., Hutton, B., and G. Dennis, "JSON Schema: A Media Type for Describing JSON Documents", Work in Progress, Internet-Draft, draft-bhutton-json-schema-01, 2020, <https://datatracker.ietf.org/doc/html/draft-bhutton-json-schema-01>.

[RFC2119]

Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, <https://www.rfc-editor.org/rfc/rfc2119>.

[RFC3986]

Berners-Lee, T., Fielding, R., and L. Masinter, "Uniform Resource Identifier (URI): Generic Syntax", STD 66, RFC 3986, DOI 10.17487/RFC3986, January 2005, <https://www.rfc-editor.org/rfc/rfc3986>.

[RFC8174]

Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017, <https://www.rfc-editor.org/rfc/rfc8174>.

[RFC9110]

Fielding, R., Ed., Nottingham, M., Ed., and J. Reschke, Ed., "HTTP Semantics", STD 97, RFC 9110, DOI 10.17487/RFC9110, June 2022, <https://www.rfc-editor.org/rfc/rfc9110>.

20.2. Informative References

[AGTP-CERT]

Hood, C., "AGTP Agent Certificate Extension", Work in Progress, Internet-Draft, draft-hood-agtp-agent-cert-00, 2026, <https://datatracker.ietf.org/doc/html/draft-hood-agtp-agent-cert-00>.

[AGTP-MERCHANT]

Hood, C., "AGTP Merchant Identity and Agentic Commerce Binding", Work in Progress, Internet-Draft, draft-hood-agtp-merchant-identity-01, 2026, <https://datatracker.ietf.org/doc/html/draft-hood-agtp-merchant-identity-01>.

[AGTP-TRUST]

Hood, C., "AGTP Trust and Verification Specification", Work in Progress, Internet-Draft, draft-hood-agtp-trust-00, 2026, <https://datatracker.ietf.org/doc/html/draft-hood-agtp-trust-00>.

[HOOD2026]

Hood, C., "Semantic Method Naming and LLM Agent Accuracy: A Controlled Benchmark of REST/CRUD versus Agentive API Interface Design", Working Paper Available by request. March 2026., 2026.

[RFC6570]

Gregorio, J., Fielding, R., Hadley, M., Nottingham, M., and D. Orchard, "URI Template", RFC 6570, DOI 10.17487/RFC6570, March 2012, <https://www.rfc-editor.org/rfc/rfc6570>.

A.1. Endpoint Authoring

A common pattern for endpoint authoring uses one declaration file per endpoint, dropped into a directory the server scans at startup:¶

• A human-editable declaration format (such as TOML, YAML, or JSON) carries the endpoint's verb, path, semantic block, schemas, errors, and handler reference.¶

• A handler function in the server's implementation language carries the business logic, with a uniform signature that takes a context object and returns a structured response or named error.¶

• The server scans the endpoints directory at startup, validates every declaration per Section 13.2, imports the referenced handler modules, and registers the valid endpoints in memory.¶

The reference Python implementation of an AGTP-API server uses TOML for endpoint declarations and a context → response | error handler signature. The exact TOML schema and handler patterns are documented at:¶

https://agtp.io/api/reference/authoring

The TOML format is one serialization of the endpoint primitive defined normatively in Section 6. Other implementations are free to use other serializations (YAML, JSON, programmatic registration, database-backed registries) provided the in-memory endpoint structure conforms to this specification.¶

A.2. Runtime Dispatch Pattern

A typical request lifecycle in a conforming server:¶

1. The server receives an AGTP request and parses the verb and path.¶

2. The dispatcher matches verb-and-path against the endpoint registry.¶

3. On match, the dispatcher validates the request body against the endpoint's input_schema.¶

4. The dispatcher checks the agent's Authority-Scope against the endpoint's required_scopes declaration.¶

5. The dispatcher constructs a context object containing validated input and agent identity, and invokes the handler.¶

6. The handler executes business logic and returns a structured response or named error.¶

7. The dispatcher translates the result into an AGTP response.¶

This pattern keeps contract enforcement in the dispatcher and business logic in the handler. Handlers do not perform schema validation, authority checks, or status code translation; the dispatcher does those uniformly across all endpoints.¶

A.3. Tooling

The reference implementation publishes tooling at agtp.io/api/tools:¶

• An OpenAPI → AGTP-API conversion tool that reads an OpenAPI 3.x specification and produces a directory of endpoint declarations with external_service handlers pointing at the original HTTP endpoints. Where method mapping requires judgment, the tool flags the endpoint for human review with suggestions from the AGTP-API approved method catalog.¶

• An AGTP-API → OpenAPI export tool for servers that need to publish HTTP-side documentation derived from their AGTP manifest.¶

• A manifest validator that checks server manifests against this specification.¶

• A conformance test runner that exercises the conformance requirements of Section 13.1 against a deployed server.¶

• An agtp-catalog-diff tool that diffs two catalog versions and optionally scans a deployment directory to identify endpoints, recipes, and policies.methods entries that would break under a candidate catalog upgrade. See Section 4.¶

These tools are non-normative reference implementations. Other parties MAY publish equivalent tools.¶

A.4. Operator Notes on Catalog Versioning

The catalog version a server declares in its manifest is the version the server validates incoming method names against. Operators upgrading the on-disk verb catalog (typically by regenerating methods.json from a new catalog release) SHOULD restart the server to pick up the new version. Long-running servers that have not been restarted after a catalog upgrade may continue validating against the previous version they loaded at startup.¶

The reference agtp-catalog-diff tool is the recommended way to preview the impact of a catalog upgrade against a deployment directory before restarting. Run it as a deployment gate to surface endpoint-removal conflicts, recipe-step references to removed methods, and policies.methods entries that would be silently skipped under the new version.¶

B.1. Wrap-and-Expose

The lowest-effort path for an organization with an existing HTTP API. The HTTP API stays where it is. The organization writes AGTP-API endpoint declarations whose handler binding is external_service (Section 12.3), pointing at the existing HTTP endpoints. AGTP-API becomes a thin agent-facing facade over the existing HTTP backend. Request translation maps AGTP-API input fields to HTTP request shapes; response translation maps HTTP responses back to AGTP-API output shapes.¶

The OpenAPI → AGTP-API conversion tool (Appendix A) produces this kind of deployment by default: it reads an OpenAPI specification and generates a directory of endpoint declarations with external_service handlers. A 200-endpoint HTTP service can be made AGTP-discoverable in a single working session.¶

Wrap-and-expose deployments SHOULD populate the manifest's apis array (Section 8) with the HTTP API descriptor so agents that need the underlying HTTP documentation can locate it.¶

B.2. Translate-and-Rehost

A medium-effort path. The organization reimplements its HTTP API as native AGTP-API endpoints, moving the handler logic into the AGTP-API server's implementation language directly. The HTTP layer goes away, or is retained only for legacy clients that have not migrated.¶

The conversion tool generates endpoint declarations and handler stubs from the existing OpenAPI specification, leaving the organization to fill in the business logic. The work is conversion, not greenfield development.¶

Translate-and-rehost deployments leave the apis array empty once the HTTP layer retires; during transition, the array continues to reference the HTTP face alongside the native AGTP endpoints.¶

B.3. Coexist-Permanently

No migration. The HTTP API stays as the primary interface; AGTP-API is added as a parallel face for agent traffic. New agent-facing features go in the AGTP-API layer; existing HTTP traffic continues unaffected. Two interfaces, one backend, indefinite coexistence.¶

This is the path most large enterprises will take when regulatory, contractual, or customer-relationship commitments make a forced migration impractical. AGTP-API is additive in this posture, not replacing.¶

Coexist-permanently deployments populate apis with HTTP API descriptors that include http_documentation URLs pointing at the HTTP-side OpenAPI specifications.¶

B.4. Path Selection

The right path depends on organizational constraints, not protocol preferences. Wrap-and-expose is the path that scales: it is the one that can move millions of existing APIs to AGTP-API discoverability at low cost. Translate-and-rehost is the right target for organizations committing to AGTP-API as the long-term primary surface. Coexist-permanently is the honest answer for organizations that want agent access without disruption to existing systems.¶