What is execution token architecture for AI agents?

Execution token architecture gives every approved AI agent action a cryptographic token (JWT) that proves the action was evaluated and permitted by the governance engine — creating an unforgeable chain of custody from decision to execution.

How execution tokens work in Exogram:

1. Agent proposes action: Sends the action payload to /v2/evaluate
2. Policy engine evaluates: All 8 gates run in 0.07ms. If all pass, an execution token is generated
3. Token contains: Action hash, state hash (SHA-256), namespace, timestamp, expiry, one-time-use claim
4. Agent commits action: Sends the token to /v2/commit. Exogram verifies the token is valid, unexpired, and the state hash still matches
5. Token consumed: After successful commit, the token is marked as used — replay attempts return 409

This architecture prevents:

• Replay attacks: Tokens are one-time-use — can't be replayed
• TOCTOU attacks: State hash verification ensures the world hasn't changed between evaluate and commit
• Token forgery: Cryptographic signing prevents agents from creating their own execution tokens
• Stale commits: Token expiry ensures actions are executed promptly or re-evaluated