Agent Definition — References

Formats for declaring an agent's surface, and the contract machinery that makes a declaration enforceable.

Declaring the callable surface

• Model Context Protocol — the specification for describing tools, resources, and prompts to a model client; the tool half of what an agent artefact has to reference.
• JSON Schema — the vocabulary most agent input and output contracts are ultimately expressed in, including the function-calling schemas accepted by the major model APIs.
• OpenAPI Specification — the long-standing prior art for separating a service's declared interface from its implementation, and the source of most of the design arguments that recur here.
• Agent2Agent (A2A) agent card — a published format for an agent's advertised identity, skills, and endpoint. Cited by name; the hosting URL has moved between organisations and no stable path is recorded here.

Making the contract enforceable

• Pydantic — runtime validation and coercion of typed models, widely used to turn a declared return shape into one that is actually checked.
• Semantic Versioning — a convention worth adopting for artefact revisions, since callers need to distinguish a wording fix from a changed return shape.

Behaviour written as rules

• Constitutional AI: Harmlessness from AI Feedback (Bai et al., 2022), arXiv:2212.08073 — an explicit, written set of principles used as the governing text for a model's behaviour, and a useful counterpoint to instructions that grow by accretion.

No verified reference is recorded for agent-artefact schemas as a standard; the formats in use are framework conventions rather than specifications, and are listed in the stack note rather than here.