When an AI agent finishes a task, I do not only want a final answer.
I want an operating record.
Here are the five receipts I want from every run.
1. Setup receipt
What agent ran? Which model/provider did it use? Which project, environment, and config were loaded? Which MCP servers or tools were available?
Without this, a successful run is hard to reproduce and a failed run is hard to debug.
2. Tool receipt
Every consequential tool call should have a compact record: tool name, normalized params or hash, result, latency, error state, and whether the call changed anything.
3. Approval receipt
If a human approved something, record what they approved. Not just "approved" in a transcript, but capability, scope, policy, timestamp, and run id.
4. Evidence receipt
If the agent made a claim or decision, what evidence did it use? File path, command output, API response, test result, or artifact.
5. Recovery receipt
If the run failed, what can be retried? What state changed? What should be rolled back or resumed?
This is the shape we are building toward with Armorer and Armorer Guard.
Armorer is the local control plane for running and supervising agents:
https://github.com/ArmorerLabs/Armorer
Armorer Guard is the runtime decision/receipt layer:
https://github.com/ArmorerLabs/Armorer-Guard
If this is a problem you feel in your agent workflows, feedback or stars on the repos would help a lot.
Top comments (0)