
Agent Audit Trails
Agents you can sign off
You cannot sign off an agent that cannot show its work. So its record is written while it runs.
- Each action written to a log kept with the work
- A named person against each consequential decision
- Refused actions recorded with their reason
- A trail your assessor reads without reconstruction
Example record
The record you hand the assessor
Appended while the agent works, with a name against each decision.
Nobody rebuilds this afterwards.
- Logged
Edited src/billing/invoice.ts
agent session, unit 3
- Refused
Read .env
secret path
- Signed
Merge to main
approved by a named reviewer
- Awaiting
Deploy to production
waiting for your release manager
The question that comes after the incident
Somebody asks what the agent did and who approved it. A chat transcript does not answer that.
Logs that keep the output and lose the steps
Approvals given in a meeting and written nowhere
Tool calls with no record of their arguments
A reconstruction that takes weeks and stays incomplete
Capabilities
What your record holds
Configured in your repository, so it holds on each machine that runs the agent.
Model
What we fix, and what you decide
What we fix
- F1An append-only action log
- F24 named sign-offs
- F3Refusals recorded with their reason
What you decide
- C1Where the log is stored and for how long
- C2Who holds each sign-off
- C3Which actions need approval first
Next step
Show us an agent you need to sign off.
We list what its record is missing and how to close it. Then you can put your name to it.

