Experimental open protocol
Governance for decisions no single agent should control.
AGP is a portable governance layer for multi-agent systems. It makes proposals, evidence, authority, ballots, resolution and history independently verifiable.
01 — Why AGP
Communication and orchestration are not governance.
Existing layers
They move work.
- MCP connects models to tools and data.
- Agent protocols enable communication.
- Workflow engines coordinate execution.
AGP
It governs authority.
- Who was authorized to decide?
- Which evidence version was reviewed?
- How did ballots resolve deterministically?
- Can an independent auditor replay the decision?
02 — Architecture
A verifiable chain from proposal to audit.
01ProposalIntent and decision scope
→
02EvidenceVersioned decision inputs
→
03BallotsSigned authority actions
→
04ResolutionDeterministic outcome
→
05AuditIndependent replay
Deterministic
Independent implementations produce byte-identical receipts from the same canonical input.
Authority-aware
Membership, roles, vetoes, revocation and quorum are explicit decision inputs.
Tamper-evident
Hash-linked history exposes omitted, altered, reordered and truncated events.
03 — Reproducible evidence
Claims backed by executable tests.
$ python3 run_benchmark_all.py
AGP v0.3 Conformance: 260/260 passed
AGP v0.4 Signed Conformance: 10/10 passed
AGP v0.5 Transparency: 8/8 passed
Workflow attack detection: 0/8
AGP attack detection: 8/8
AGP BENCHMARK COMPLETEThe workflow baseline intentionally trusts coordinator state. The benchmark demonstrates AGP's explicit verification model; it is not a universal claim that all workflow engines are insecure.
04 — Project status
Experimental by design. Honest about its limits.
Implemented
- Python and Go resolvers
- Signed envelopes
- Revocation and replay controls
- Append-only transparency log
- Automated GitHub conformance
Next
- Independent reproduction
- Third implementation
- Formal schemas
- Real-world pilot
- External security review
Not claimed
- Production maturity
- Evidence truthfulness
- Confidentiality
- Global consensus
- Legal enforceability
Do high-impact multi-agent systems need a governance layer distinct from communication and orchestration?
AGP exists to test that hypothesis in public.
Review, reproduce or challenge AGPPublic Demo
Understand AGP in 120 seconds
Run a dependency-free scenario where a coordinator changes a deployment payload after approval. A conventional workflow accepts it, while AGP detects the mismatch and reproduces the decision independently.
python3 examples/120_second_demo/demo.py
PUBLIC REVIEW DRAFT
AGP Whitepaper v0.9.2
Read the technical foundation of the Agent Governance Protocol, including its architecture, deterministic resolution model, threat model, conformance results, limitations, and research roadmap.