I'm building AI agents
Learn which released specifications you can implement, how to pin a version, and how to describe the limits of your implementation.
Start building →
Governance Commons · Open AI governance specifications
Who authorized an AI agent to act—and how can you verify what happened?
Governance Commons publishes open specifications for describing agent identity, authority, and coordination, with evidence that people can inspect. Start with an explanation, then explore the technical details at your own pace.
A maintainer-led commons. Not a regulator or certification authority.
Start here
What brings you to Governance Commons?
Choose a path. You do not need to know the component names to begin.
I'm reviewing or auditing
Understand inspectable governance evidence and what a validation result does—and does not—establish.
Review evidence →I'm researching governance
Explore the sociotechnical questions, component boundaries, and published status of the work.
Explore the ideas →I want to contribute
Read the public source, examine open design questions, and offer technical critique without assuming a formal membership program exists.
Explore the repositories →How it fits together
Describe the agent. Establish the boundaries. Preserve the evidence.
Governance Commons supplies specifications and evidence contracts—not an agent execution engine or a grant of authority. A real system must establish legitimate authority through its own people and institutions.
Describe identity and context
Use Agent Dossier to explore structured descriptions of agents and their governance context.
Inspect coordination boundaries
Explore Agent Matrix for the relationships and governance structure relevant to multi-agent work.
Keep evidence reviewable
Use the conformance guidance to understand verification records, supported checks, and claim limits.
Across these concerns, Ontic Namespace Structure (ONS) provides a naming structure. Human approval, review, override, and recourse remain governance concerns—not capabilities automatically supplied by a specification.
Example: If one AI agent delegates work to another, ask who each agent is, what authority the actual system recognizes, what coordination rules apply, and what evidence a human can inspect afterward. The specifications help structure those questions; they do not authorize the delegation themselves.
Current public state
RFC-0001 is accepted; its new public boundary is not yet active.
The accepted Public Boundary Model separates the umbrella, independently adoptable specifications, reference tooling, publication infrastructure, and maintainer workspace. Activation still depends on gates G-1 through G-5. GC v1.2.0 remains the current pinned release record while that transition is verified.
The sociotechnical problem
Capability is not the same as legitimate authority.
Governed systems need more than schemas or agent routing. They need inspectable answers about identity, authority, delegation, human oversight, handoffs, provenance, evidence, accountability, institutions, trust, legitimacy, power, appeal, override, revocation, and human–agent coordination.
Authority before action
Record who may act, under what scope, from which source of authority, and how that authority expires or is revoked.
Evidence after action
Preserve validation, handoff, provenance, telemetry, and decision evidence without pretending that evidence itself grants authority.
Human recourse
Keep approval, refusal, appeal, override, escalation, and accountability visible when work crosses human and institutional boundaries.
Accepted boundary model
Five public roles, not five equal standards.
This is the transition authorized by RFC-0001. It becomes active doctrine only after all activation gates verify.
Public umbrella
Mission, shared vocabulary, publication identity, and community-facing orientation. Citeable, not adoptable.
Specifications
Agent Dossier, Agent Matrix, and ONS are independently versioned adoption units.
Reference tooling
Validators, SDKs, CLIs, evidence utilities, and runtime candidates implement selected contracts.
Publication infrastructure
Websites and repositories publish source, releases, documentation, schemas, and status.
Maintainer workspace
Plans, release records, internal coordination, and construction history are not products.
Public components
Role, maturity, version, and baseline status stay visible.
Released specificationsReleased
Agent Dossier
Per-agent governance specification
Per-agent identity, authority, handoff, telemetry, and audit contracts.
v1.1.0 · Pinned in GC v1.2.0
Released specificationsReleased
Agent Matrix
Multi-agent governance specification
Multi-agent capability, routing, trust, safety, and coordination.
v1.4.0 · Pinned in GC v1.2.0
Released specificationsReleased
Ontic Namespace Structure
Naming and namespace specification
Naming grammar, namespace identity, collision rules, and validation IDs.
v1.3.0 · Pinned in GC v1.2.0
Reference runtimeCandidate
Agent Project Orchestrator
Execution runtime candidate
Project-level task graphs, dispatch, handoffs, recovery, telemetry, and audit.
v1.1.0 · Historically pinned in GC v1.2.0
Emerging workPre-release
Agent Team Protocol
Fleet delegation governance specification
Delegation, trust topology, shared audit, escalation, and team lifecycle.
v0.1.0 · Not in GC v1.2.0
Evidence contractsEvolving contract
Governance Record
Durable governance evidence contract
Durable evidence for authority, approval, action, handoff, lifecycle, and provenance.
v1.0.0 · Implemented in Python toolkit 0.2.0
Evidence contractsImplemented contract
ConformanceReport
Validation result contract
Machine-readable rule results, summary counts, target version, profile, and overall outcome.
v1.0.0 · Shared validator output format
Reference toolingAlpha
GC Toolkit / Registry
SDK, validators, CLIs, and evidence utilities
Reference implementation for supported validation, conformance reporting, discovery, and evidence contracts.
v0.2.0 · Python 0.2.0 published; npm 0.2.0 pending
Adopt
Make a narrow claim that another person can inspect.
Name
Choose one released specification and pin the exact version you implement.
Verify
Use a published validator only where it supports that target. Otherwise document manual checks and label the result provisional.
Preserve
Retain the validator version, ConformanceReport or manual evidence, source provenance, and limits of the claim.
Research, education, and community
The public purpose is broader than the repositories.
Today, GC publishes specifications, evidence contracts, reference tooling, an ecosystem map, and maintainer-led explanatory material. The goals below are directions for a commons—not claims that formal programs or neutral institutions already exist.
Explore community & collaboration →
Research
Connect technical mechanisms to sociology, institutions, legitimacy, power, trust, accountability, and the real effects of delegated machine action.
Education
Develop approachable examples, curricula, case studies, and critical vocabulary for builders, reviewers, organizations, and affected communities.
Community
Invite critique, implementation experience, competing perspectives, and future shared stewardship without claiming neutrality or participation that does not yet exist.
Authority boundary
A commons, not a regulator or certifier.
Governance Commons is not ISO, NIST, IEEE, AAIF, a regulator, or a certification authority. External organizations retain authority over their own work; GC must preserve their provenance and state relationships as alignment, mapping, implementation, or research—not endorsement.