agent-manifest

Agent Manifest — Specification (v0.1.0)

Agent Manifest is a minimal, execution-agnostic, declarative specification that allows AI agents to publicly declare identity, purpose, constraints, autonomy, risk boundaries, stopping authority, audit posture, and data handling before interaction.

This specification declares. It does not execute, validate, score, enforce, or decide.

---

1) Required fields

An Agent Manifest MUST declare the following fields:

---

Identity

• manifest_version (must be 0.1.0)
• agent_id (stable identifier)
• agent_name (human-readable)
• agent_version (agent’s own version)

---

Ownership and accountability

• owner.type (individual

  organization

  system)

• owner.identifier (accountable reference)

Agents without a declared owner are invalid by design.

---

Purpose

• purpose.primary_code (bounded code)
• purpose.description (specific, bounded explanation)

---

Hard constraints

• forbidden_actions[] (minimum 1 entry)

What is not declared is considered prohibited.

---

Autonomy

• autonomy.level (integer 0–3)

Autonomy describes decision freedom, not intelligence.

Autonomy Levels (normative interpretation)

autonomy.level is a declared execution posture (not a capability claim):

• 0 — No execution: advisory / descriptive only
• 1 — Supervised: actions only under explicit user instruction or approval
• 2 — Scoped autonomy: bounded workflows within declared constraints
• 3 — High autonomy: can execute within declared constraints without per-step approval

---

Risk profile

• risk_profile.level (low

  medium

  high)

• Optional: risk_profile.notes

Risk is self-declared. This specification does not evaluate.

---

Stopping authority

• stopping_authority.stoppable_by[]
• stopping_authority.mechanism
• Optional: stopping_authority.stages[]

Every autonomous system must be stoppable. Agents without declared stopping authority are structurally non-compliant.

---

Audit surface

• audit_surface.logging (none

  basic

  detailed)

• audit_surface.reconstructability (none

  partial

  full)

• audit_surface.opacity_declared (boolean)
• Optional: audit_surface.notes

Accountability requires reconstructability. Opacity, when present, must be explicitly declared.

---

Data handling

• data_handling.stores_personal_data (boolean)
• Optional: data_handling.retention

If personal data is stored, retention must be declared.

---

Contact

• contact.email

---

2) Optional: Language

• language.primary
• language.supported[]

This is a human-facing declaration. Inter-agent communication is assumed to use structured data.

---

3) Extensibility

This specification is intentionally extensible.

Implementers MAY include additional fields without breaking compatibility.

To preserve interoperability, custom or domain-specific fields SHOULD be placed under an extensions object.

---

4) Schema

Validation is performed against:

• spec/manifest.schema.json

---

5) Versioning

This specification follows Semantic Versioning.

Current version: 0.1.0.

This site is open source. Improve this page.