What did this agent commit to,
and when was it recorded?

A design pattern for accountable autonomous agents

Agents act. When something goes wrong, one question follows — and today it is almost always unanswerable. Logs tell you what the agent says it did. Prove Before Act makes it commit before acting.

Example — a staging deployment. WHY: the requested fix is ready. Verified commitment: the agent commits to deploy the reviewed release to staging. Independent proof establishes the commitment’s hash and its order in the chain. Operator-permitted action: the operator configures the deployment gate to allow that deployment only after the commitment is recorded. WHAT: a later record can state the deployment result. This proves what was committed and when—not that the off-chain deployment happened, or that any arbitrary off-chain action occurred.

T(intent_proof) < T(action)
Action timing requires an operator gate or independent witness.