What stops an AI agent from acting without proof? One rule: no proof, no action.
AGP is the rulebook that makes every AI action leave a signed record behind.
AGP is the rulebook behind KHEPRA. Every step an AI agent takes, called a Governed State Transition, must leave a signed record called an Agent Evidence Object. If it can't prove what happened, the action never happens. No exceptions, no defaults to yes.
What rules does AGP never bend?
Trust is not enough. Bitcoin proved this with money. AGP does the same for AI actions.
No agent gets full control. Every permission is limited and tracked.
If anything is unclear, the action is blocked. There is no default yes.
Every record is built the same exact way, every time. Anyone can check it and get the same result.
You don't have to trust us. You can run your own checker and replay any chain of proof yourself.
Locked with ML-DSA-65, a signature future quantum computers can't crack. Ready to upgrade as standards change.
What does a signed record actually look like?
Every action an agent takes creates exactly one record. It's locked to the record before it and signed with ML-DSA-65, so no one can slip an action past it.
{
"aeo_id": "b3:hash(canonical_bytes)",
"previous_hash": "b3:parent_aeo_id",
"timestamp": "2026-02-17T10:30:00Z",
"gst_phase": ["Intent", "Authorization", "Actuation", "Verification", "Attestation"],
"agent_id": "ml-dsa-65:did:khepra:agent/finance-copilot#v3",
"human_approver": "ml-dsa-65:did:khepra:user/j.okafor",
"intent": {
"mission_id": "generate_q3_ap_summary",
"desired_state": { "report": "q3-ap.parquet", "classification": "CUI" }
},
"policy_applied": ["CMMC_L3_SC.3.177", "NIST_800-53_AC-2"],
"privilege_context": { "symbol": "Eban", "scope": "warehouse.reports" },
"execution": {
"command": "warehouse.query(...)",
"timestamp": "2026-02-17T10:30:00.412Z"
},
"verification": {
"pre_state": "b3:9c3a...",
"post_state": "b3:1f8e...",
"equality": true
},
"signature_ml_dsa_65": "base64(...)",
"governance_graph_edges": [
"agent_id→mission_id",
"policy_id→execution_id"
]
}What stands between an agent's idea and a real action?
Seven checks. Every one has to pass before the next one can start.
The agent states what it wants to do, and that gets signed first.
The request is checked against the rules. Extra conditions get added if needed.
ASAF hands out a limited permission just for this one action. It can be taken back.
ASAF runs the action carefully. If a check fails, it undoes the action right away.
The system checks the result matches exactly what was promised.
The whole action gets signed and locked with ML-DSA-65.
The signed record is saved for good. It can never be quietly erased.
Questions people ask before they buy
What is the Autonomous Governance Protocol (AGP)?
AGP is the rulebook behind KHEPRA. Every step an AI agent takes, called a Governed State Transition, must leave a signed record called an Agent Evidence Object. If it can't prove what happened, the action never happens.
What is a Governed State Transition (GST)?
A GST is the seven-step path every AI action takes: Intent, Policy, Privilege, Actuation, Verification, Attestation, Evidence. Each step is checked before the next one can start.
What is an Agent Evidence Object (AEO)?
An AEO is one signed record for one action. It's locked to the record before it and signed with ML-DSA-65, so no one can slip an action past it or quietly erase it later.
Why does AGP use post-quantum signatures?
AGP signs every record with ML-DSA-65, a signature future quantum computers can't crack. It's built to upgrade as post-quantum standards change, so proof made today still holds up years from now.
Can I verify AGP proof myself, without trusting KHEPRA?
Yes. Every record is built the same exact way every time, so anyone can run their own checker and replay any chain of proof and get the same result KHEPRA got.
Last updated August 2026