FAFA: A Declarative Agent Capability Format

1.1. Problem Statement

Agent-interaction work consistently defines an object model for agent identity, capability, and endpoint, and then defers the serialization to "some JSON" or an unspecified schema. The object model is specified repeatedly; the on-the-wire, in-the-registry, in-the-repository format — the document an agent publishes about itself — is left unspecified.¶

In production that identity is usually inferred from system prompts, product settings, and protocol-native cards that do not travel cleanly across hosts. The result is seam rewrites when an agent moves between environments, repeated inference cost, and trust surfaces that cannot be diffed or reviewed as data.¶

1.2. Solution Overview

.fafa specifies the missing document: a vendor-neutral, YAML-native declaration of what an agent is, what it can do, and how it is reached. It functions as a portable passport that answers four questions:¶

• who the agent is,¶

• what it may do,¶

• where it is reached, and¶

• what it must never do.¶

A .fafa document describes an agent; it never instructs, authorizes, or executes one. It defines the document that protocols and hosts carry, reference, or resolve; it does not define discovery, transport, authorization, or orchestration.¶

.fafa is complementary by design. It composes with existing and future agent-interaction protocols and repository instruction conventions and competes with none of them. Protocol-native cards (A2A Agent Cards, MCP Server Cards, and similar) remain useful wire formats; files such as AGENTS.md remain the repository ops briefing. A .fafa document is the house-neutral source of truth that can be projected into those formats (see Section 4).¶

A companion paper provides extended motivation, worked examples, and further detail on the projection model summarized in Section 4 [FAFA-PAPER].¶

1.3. Design Goals

1. Declarative, never executable — a .fafa document describes an agent; it never instructs, authorizes, or executes one.¶

2. Vendor-neutral — no dependency on any single agent-interaction protocol.¶

3. Composable — sits at the format layer beneath protocols, not in competition with them.¶

4. Forward-compatible — unknown fields are preserved as extensions; documents remain valid across minor revisions.¶

5. Family-coherent — extends .faf [FAF-FORMAT]; shares YAML 1.2 [RFC9512] as base syntax and MIT-licensed distribution.¶

6. Portable source of truth — intended to be authored once and projected into the protocol- or host-specific cards that different environments require.¶

3.1. Format Overview

A .fafa file MUST be a valid YAML 1.2 document [RFC9512] [YAML]. It is declarative: it carries identity, capability declarations, and endpoint references, and MUST NOT contain executable content, code, or instructions to a model.¶

3.2. Media Type

The agent format is identified by application/vnd.fafa+yaml, registered in the IANA vendor tree on 2026-06-26 [IANA-FAFA]. This document documents that vendor-tree registration. No standards-tree registration is requested. The canonical specification is [AGENT-FORMAT]. The file extension is .fafa.¶

3.3. Document Structure

A .fafa document is a single YAML mapping.¶

Required top-level fields: version; agent (with at least name and id); capabilities (array, MAY be empty); endpoints (array, MUST be non-empty).¶

Optional top-level fields: attachment, provenance, signature, metadata (free-form vendor extensions).¶

Primary example:¶

version: "1.0"

agent:                       # REQUIRED -- identity
  name: faf-agent
  id: "did:web:faf.one:agent"
  vendor: WolfeJAM
  version: "1.0.0"
  description: "Cites the spec or refuses out of scope"

capabilities:                # REQUIRED -- MAY be empty
  - name: validate
    type: tool
    description: "Validate .faf documents against the specification"
    tags: [validation, faf, spec]
    cites_spec: "application/vnd.faf+yaml"

  - name: score
    type: tool
    description: "Compute AI-readiness score for a .faf document"
    tags: [scoring, faf]

  - name: answer
    type: tool
    description: "Answer questions about the FAF family of formats"
    tags: [documentation, faf]
    apophatic: true
    cites_spec: "application/vnd.faf+yaml"

endpoints:                   # REQUIRED -- MUST be non-empty
  - protocol: mcp
    transport: stdio
    location: faf-agent-mcp

  - protocol: mcp
    transport: http
    location: "https://mcp.faf.one/faf-agent"

provenance:                  # OPTIONAL -- context substrate
  faf: "https://faf.one/.well-known/faf"
  spec: "application/vnd.faf+yaml"

Minimal valid example:¶

version: "1.0"

agent:
  name: example-agent
  id: "did:web:example.com:agent"

capabilities: []

endpoints:
  - protocol: mcp
    transport: stdio
    location: example-agent

3.4. Agent Section

Declares identity.¶

• Required: name (string, unique within the issuing vendor namespace); id (globally unique identifier — DID, URI, or vendor-scoped UUID).¶

• Optional: vendor; version (semantic version of the agent, NOT of this specification); description; homepage; license (SPDX identifier).¶

3.5. Capabilities Section

Each capability is a structured object.¶

• Required: name (unique within the document); type (tool, resource, prompt, event, or other).¶

• Optional: description, input_schema, output_schema, tags, apophatic (boolean; if true, the capability refuses out-of-scope invocation rather than attempting best effort), cites_spec (the format or specification the capability operates against).¶

The capabilities array MAY be empty. Declaration is not authorization (see Section 5).¶

3.6. Endpoints Section

Each endpoint declares a protocol binding.¶

• Required: protocol (mcp, a2a, grpc, http, or other); transport (stdio, http, websocket, quic); location (reachable address — URI, package name, or transport-specific locator).¶

• Optional: version, auth, region, health.¶

Multiple endpoints MAY declare the same protocol with different transports.¶

3.7. Attachment Section (Optional)

Describes the infrastructure context within which the agent operates, without requiring any particular gateway protocol. Optional fields: gateway, domain, policy, attachment_point.¶

3.8. Provenance Section (Optional)

Cites the context substrate the agent operates from — the FAF-family integration point. A .fafa agent SHOULD declare which .faf it reads. Optional fields: faf (reference to a .faf document), fafm (reference to a .fafm document), spec (media type of the substrate), version.¶

3.9. Parser Requirements

• Consumers MUST use YAML safe-load (no custom type construction).¶

• Consumers MUST treat unknown top-level fields, and unknown subfields within agent, capabilities, and endpoints, as forward-compatible extensions.¶

• Consumers SHOULD use UTF-8 encoding and SHOULD keep documents within common registry storage budgets (RECOMMENDED: < 64 KB).¶

3.10. File Conventions

• Standard filename: agent.fafa (alongside project.faf at the project layer), or served via the /.well-known/faf discovery anchor defined in [FAF-FORMAT].¶

• File extension: .fafa. Encoding: UTF-8.¶

• Multiple .fafa documents per scope are permitted, distinguished by agent.id.¶

3.11. Versioning

The top-level version field declares the version of this specification to which the document conforms. Versioning is semantic: 1.x revisions are backward-compatible; a new major version is reserved for breaking changes made only with strong cause. Consumers MUST accept documents whose version shares their supported major version, and MUST apply forward-compatible extension handling for unknown fields. This separates the agent's version (agent.version) from the format's version (top-level version), so an agent may revise independently of the specification.¶

4.1. Projection, Not Replacement

.fafa is not a replacement for protocol-native identity and capability documents; it is a source of truth those documents MAY be projected from. Existing wire formats remain useful and unaffected:¶

The preferred pattern is one .fafa passport with many projections, rather than multiple hand-maintained, independently drifting cards describing the same agent. A publisher MAY maintain .fafa as the authoring source and generate protocol-native documents from it; this document does not require or standardize that generation process.¶

4.2. A2A Agent Cards

The Agent-to-Agent (A2A) protocol defines an Agent Card as a JSON metadata document, conventionally published at /.well-known/agent-card.json. An A2A Agent Card describes identity, skills, supported interfaces, security schemes, and related metadata, and MAY be signed.¶

A .fafa document can be projected into an A2A Agent Card. Typical mappings include agent.name / agent.id → identity fields, capabilities[] → skills, endpoints[] → supportedInterfaces, and agent.description → card description. The .fafa document remains the upstream source; the Agent Card is the protocol-specific advertisement derived from it.¶

4.3. MCP Server Cards and Tool Surfaces

The Model Context Protocol (MCP) is evolving discovery mechanisms, including Server Cards that describe a server's identity, transports, and protocol versions prior to connection. Tool, resource, and prompt surfaces remain runtime concerns (for example via tools/list).¶

A .fafa document can serve as the upstream source that is projected into an MCP Server Card or used to seed the tool surface an MCP host advertises; .fafa does not replace MCP Server Cards or the runtime tool list.¶

4.4. Repository Instruction Files

Files such as AGENTS.md, CLAUDE.md, GEMINI.md, and .cursorrules provide project-level operational guidance. They are instruction surfaces, not structured identity or capability declarations.¶

.fafa and these files address orthogonal questions: .fafa answers who the agent is and what it may do; AGENTS.md answers how to work in this repository. The two coexist cleanly.¶

4.5. Relationship within the FAF Family

.fafa is the agent member of the FAF family [FAF-FORMAT]:¶

.faf   -> Context Layer   (project IS -- static, read once)
.fafm  -> Memory Layer    (agent REMEMBERS -- mutating)
.fafa  -> Agent Format    (agent IS -- declarative identity)

The .fafa schema extends application/vnd.faf+yaml additively. A consumer that understands only .faf parses a .fafa file as valid YAML and extracts top-level identity; a .fafa-aware consumer gains the full agent declaration.¶

5.1. Active or Executable Content

None. A .fafa document is YAML data with no scripts, code, or active markup. Implementations MUST treat .fafa content as data, never as instructions or commands. Declaration is not authorization: an implementation MUST NOT invoke or auto-execute a capability solely because it is declared.¶

Projection of a .fafa document into a signed protocol card does not transfer authorization; the receiving system must still apply its own policy and least-privilege controls.¶

5.2. YAML-Specific Vulnerabilities

.fafa is built on YAML 1.2; the YAML security considerations of [RFC9512] apply. Parsers MUST disable custom type construction (e.g., !!python/object, !ruby/object), MUST enforce alias-expansion limits to mitigate entity-expansion ("billion laughs") attacks, and MUST enforce nesting-depth limits to prevent stack overflow.¶

5.3. Untrusted Input and Prompt Injection

.fafa documents originate from agents and third parties and MUST be treated as untrusted input. Consumers MUST NOT elevate .fafa content to "system" or "developer" tiers within prompt hierarchies; it occupies the user-data tier. Capability declarations are a documented vector for over-trust and prompt injection.¶

5.4. Identity, Endpoints, and Least Privilege

Identity claims, capability declarations, and endpoints are self-asserted. Consumers MUST treat endpoints as untrusted until verified (against signature where present), MUST scope capability consumption to the declaring agent's verified identity, and MUST apply least privilege when acting on declared capabilities.¶

5.5. Prohibited Content

.fafa documents MUST NOT contain secrets, API keys, or credentials; MUST NOT contain user-personal data (use .fafm for memory of that kind); and MUST NOT contain executable code.¶

6.1. Media Type (Documentation Only)

The agent format is already registered in the IANA "Media Types" registry, vendor tree:¶

• application/vnd.fafa+yaml — registered 2026-06-26 [IANA-FAFA]. File extension .fafa. Published specification: [AGENT-FORMAT].¶

This document documents that registration. No new media-type registration is requested.¶

6.2. Relationship to the Existing Vendor-Tree Registration

The vendor-tree registration remains valid. This document does not request promotion to the standards tree.¶

7.1. Normative References

[AGENT-FORMAT]

Wolfe, J., ".fafa — FAF Agent Format Specification, Version 1.0", August 2026, <https://github.com/Wolfe-Jam/faf/blob/main/AGENT-FORMAT.md>.

[IANA-FAFA]

IANA, "Media Type: application/vnd.fafa+yaml", June 2026, <https://www.iana.org/assignments/media-types/application/vnd.fafa+yaml>.

[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>.

[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>.

[RFC9512]

Polli, R., Wilde, E., and E. Aro, "YAML Media Type", RFC 9512, DOI 10.17487/RFC9512, February 2024, <https://www.rfc-editor.org/rfc/rfc9512>.

[YAML]

Ben-Kiki, O., Evans, C., and I. döt Net, "YAML Ain't Markup Language Version 1.2", October 2009, <https://yaml.org/spec/1.2/>.

7.2. Informative References

[FAF-FORMAT]

Wolfe, J., "FAF: YAML Media Types for AI Project Context and Persistent Memory", Work in Progress, Internet-Draft, draft-wolfe-faf-format, 2026, <https://datatracker.ietf.org/doc/html/draft-wolfe-faf-format>.

[FAFA-PAPER]

Wolfe, J., "Why Agents Need a Passport: .fafa — Portable Identity for the Agentic Era", Zenodo, DOI 10.5281/zenodo.21951641, 15 August 2026, <https://doi.org/10.5281/zenodo.21951641>.