What should an AI-agent incident record contain after a bad output is already sent?

The bad output is already sent; the next failures are usually record failures. Teams reconstruct from memory days later, correct the recipient without documenting what was wrong, and change three settings at once so nobody knows which one mattered. The incident record prevents all three by fixing the shape of the account while the facts are still in the systems.

The backbone is the work record itself. The instruction and the steps, with their saved results, show what the agent was asked and what it did, including the step where the output went wrong and the approval state it passed through. Where a step was accepted by a human, the record says by whom; where no human saw it, the record shows that too, which is the finding the control change must answer.

Write the recipient section with particular care: the exact output as sent, the channel, the time, and who received it, taken from the sending system rather than summarized. Then the correction: what was sent to repair it, when, and its record. Close with one control change, named and dated: a human sign-off added before sends, a spend or action limit tightened, a source check made mandatory. One change, owned by a person, reviewed on a date. The template below is the whole file; incidents kept this way become the cheapest reliability programme a small team can run.

Worked example

Item 2 is the section that exists only if the work was recorded: on a recorded job, the step list and approval states read like the public record at https://zambo.dev/handover/bd7e5325-33fd-4c5e-b029-76b781fa2f51 , where five steps and the acceptance state are visible on one object. On an unrecorded job, item 2 becomes a reconstruction, and the incident record should say so.

Frequently asked questions

How fast should the incident record be written?

The skeleton (items 1 to 5) the day you discover the incident, while systems still hold the facts. Root cause and the control change can follow within days, but a skeleton written late is a reconstruction, and the record should label itself honestly either way.

Should the incident record be shared with the client?

The exposure and correction sections usually should, promptly and factually; clients forgive errors faster than they forgive discovering them alone. The full internal record, including root cause debate, is yours; what you send is a decision for your contracts and your judgement.

What is the most common missing item?

The approval state of the bad step. Teams can show what the agent produced but not whether anyone saw it before it sent. If your records cannot answer that, the control change is already chosen for you: recorded sign-off before sends.

Try a handover

A handover gives your client one proof link with verifiable receipts to review, then Accept or Flag.

try it: say 'use zambo to hand this over to <client>' in your ai

Add Zambo · Handover docs · Real public handover · AER-1