The incident evidence set, field by field
After a costly agent mistake, the first records request follows a predictable shape, whether it comes from an insurer, a client, or your own review. The set that answers it has eight parts, and every part is far easier to keep during the run than to rebuild after it:
- The instructions. What the agent was told to do, saved as given, including the source material it was told to work from.
- The inputs and sources. What the agent actually read: the pages, files, and records behind the run, with their states at the time.
- The steps. Every tool call in order, each with its own receipt: what ran, what went in, what came back, and when. Attempts count too: a rejected call still mints a receipt, so the set shows what the agent tried, not only what it managed.
- The spend and action state. What the run cost, what limits were in force, and which actions completed before anyone stopped the run.
- The reviewer decision. Who saw the work before it went out, what they accepted or flagged, and when. If nobody reviewed it, the record shows that plainly, and that plain fact is itself part of the set.
- The customer impact. What reached the customer: the output as sent, the recipients, and the moment it landed.
- The timeline. Discovery, containment, correction, and customer notice, each tied to a time and to the record kept of it.
- The correction. What was sent or done to put it right, kept as its own record, linked to the incident rather than overwriting it.
The containment half of this story, freezing the session while it is still live, is the unauthorized action guide. The liability framing behind these questions is in the AI agent liability guide.
The record is evidence, not a verdict
Keep two things separate in your own head and in anything you send. The record shows what was saved and proves the saved result was not changed since it was saved. The claim decision, what a policy covers and what it pays, belongs to the insurer and the policy. A clean record does not decide a claim, and a messy one does not either. What the record does is end the argument about what the record says, so the decision can be made on a record anyone can recheck: an agent receipt is a record of one ai agent tool call that anyone can check: it keeps the exact saved bytes and their sha-256 digest, and it proves the result was not changed since it was saved, not that it was right. Any receipt in the set can be rechecked at zambo.dev/verify with no account.
This page is general information about records, not advice about a specific policy or claim.
Build the set before the incident, not after
- Run the work on the record. Calls routed through Zambo mint their receipts as the work happens, so the steps, the spend state, and the attempts exist before anything goes wrong.
- Save instructions as records. An instruction kept in a scrolling chat is an instruction you may not be able to produce. Save it with the workflow it belongs to.
- Record the review. A handover closed with an Accept or a Flag is the reviewer decision, already written. The pattern is the human approval record.
- Keep the fences visible. The spend cap and the input screening on the run are part of the set: they show the boundaries the agent worked inside. The checklist is how to make your AI agent safer.
Set it up on the next run
- Route the agent through the one endpoint. Connect the AI to https://zambo.dev/mcp so every call lands on the record. Start at zambo.dev/add: 20 free calls per day, no install.
- One workflow per job, so one incident maps to one closed, checkable record.
- Set the spend limit at the start of the run and shield the inputs the session will read.
- Close every job with a handover, so the reviewer decision exists for routine work too, not only for incidents.
Build the incident set before you need it: add Zambo to your AI and the record starts with the first call. Start with 20 free calls per day; the Day Pass is $0.99 USDC on Base for 24 hours of full use, and the Zambo Pass is $29 for 30 days, one payment.
Frequently asked questions
What records would an insurer ask for after an AI agent makes a costly mistake?
Eight parts: the instructions the agent was given, the inputs and sources it worked from, every step in order with its own receipt, the spend and action state including the limits in force, the reviewer decision or the plain fact that nobody reviewed it, the customer impact, the timeline of discovery and correction, and the correction record itself.
Can receipts prove the agent's work was correct?
No, and the set does not claim that. A receipt proves the saved result was not changed since it was saved. Correctness is judged from the instructions, the inputs, and the review record. The receipts end the argument about what the record says, so the judgment happens on facts anyone can recheck.
What if nobody reviewed the agent's work before it went out?
Say so, and show it. A record with no approval step is still a record, and presenting it plainly is stronger than a reconstructed review nobody can check. Then fix the routine: close each job with a handover so the next run has a recorded Accept or Flag before delivery.
How long after the incident can this set be assembled?
The receipts, caps, and approval states exist from the moment the work ran, if the run was routed through Zambo. Assembling the set is then an afternoon of gathering and ordering. Work that ran outside the connection has to be reconstructed from whatever survives, which is why the setup belongs before the incident.
What is an agent receipt?
An agent receipt is a record of one ai agent tool call that anyone can check: it keeps the exact saved bytes and their sha-256 digest, and it proves the result was not changed since it was saved, not that it was right.