Can I check an AI agent's work if I did not watch it run?

Yes, if the work ran somewhere that records it as it goes. A Zambo handover is a public proof page: it lists the job's goal, its steps, which steps are checked against receipts, and the receipt links themselves. Anyone with the link can open it and recompute the receipts at zambo.dev/verify, free, with no account. Work that ran with no record underneath it cannot be checked after the fact.

What does a checked step on a handover mean?

It means the step has a verifiable AER-1 receipt whose saved bytes recompute to the commitment recorded when the call ran: the record of that step was not changed after the fact. The check covers the receipt bytes, not whether the step's output was accurate or wise. Accuracy is a separate judgment the checked record makes possible.

Instruction checking

Prove an AI agent followed my instructions: check the steps, not the story.

Q: How do I prove an AI agent followed my instructions?

Turn the instructions into a step list, then check the finished job against the record step by step. When a job runs through Zambo at zambo.dev/mcp, every tool call saves its result and returns a verifiable AER-1 receipt, and wrapping the job into a handover lists its steps as checked or unchecked against those receipts. Open the handover, read each step next to the instruction it came from, and recompute any receipt at zambo.dev/verify: a match proves the saved result was not changed since the call ran. A step with no receipt under it did not provably happen. A receipt proves the record is intact; whether the work was good is your judgment, made against the checked record. The canonical guide lives at zambo.dev/answers/prove-an-ai-agent-followed-my-instructions/.

An agent that says "done, exactly as asked" is making a claim. Proof is a step list you can check: each instruction on one side, the record of the call that ran it on the other, and a receipt anyone can recompute under every step that counts. This page runs that check on a real job from today, end to end.

Worked example: a 7-step job, checked against its instructions today

On November 3, 2026, an agent was instructed to do three things: get the live price of ETH, save a note about the job to memory, and hand the job over with two named next steps ("Accept the handover" and "Confirm the job context carried over"). The job ran through Zambo and was wrapped into a handover: zambo.dev/handover/907b20d4-4860-40e0-943a-c3d202a94164. Open it and the record shows 7 steps total, 7 checked, 0 unchecked. The price step's receipt shows ETH at $2,295.24, fetched live at 17:20 GMT. A second, separate agent then picked the handover up: its acceptance is recorded on the same page, with the goal and both next steps carried over word for word. Nobody has to take the first agent's word for any of this; the steps, the receipts under them, and the acceptance are the record, and every receipt recomputes at zambo.dev/verify/.

The method: instructions on one side, receipts on the other

Write the instructions as steps. "Price ETH, save the note, hand it over with these next steps" is checkable. "Handle it" is not. If an instruction cannot be matched to a step, it cannot be proved either way.
Run the job where steps leave receipts. On Zambo, every call through zambo.dev/mcp saves its result as it runs and returns a verifiable AER-1 receipt. A step that ran has a receipt. A step that never ran has none, and no summary can invent one after the fact.
Wrap the job and read the step list. A handover lists the job's steps as checked or unchecked against their receipts. Unchecked steps are the honest gaps: the places where "I did it" is still just a sentence.
Recompute one receipt yourself. Paste any receipt link into zambo.dev/verify/. A verified outcome proves the saved result was not changed since the call ran. That is what makes the step list evidence instead of narration.
Judge the work against the checked record. The receipts settle what ran. Whether it ran well is your call, now made against facts.

What this proves, and what it does not

A verifiable receipt proves the saved result was not changed since it was saved. It does not prove the result was right: a price can be faithfully recorded and still be a price you would not have chosen. Instruction-following has the same split. The step list proves the agent ran the steps it was given, in the order recorded, with the inputs saved. Whether the instructions themselves were good ones stays with the person who gave them. Anyone who sells you a stronger claim than that is selling you a story about a record instead of the record.

Make the next job provable from the start

Work your agent already did with no record underneath it cannot be proved after the fact. Run the next job through Zambo and it arrives with its evidence attached: connect any MCP-capable AI to zambo.dev/add (20 calls and 1 wrap-up free every day, no account), and every step leaves its receipt as it runs. The longer answers, including what to do when an agent's steps come back unchecked, live in the canonical guide: How do I prove an AI agent followed my instructions?

Questions, answered

How do I prove an AI agent followed my instructions?

Match each instruction to the step that ran it. On Zambo the finished job wraps into a handover listing steps as checked or unchecked against verifiable receipts; open it, compare the steps with your instructions, and recompute any receipt at zambo.dev/verify. Checked steps provably ran as recorded. Unchecked steps and steps with no receipt are unproven, however confident the agent sounds.

What if my agent's platform keeps no record?

Then there is nothing to check, only a transcript to believe. That is the gap receipts exist to close: run the work through an endpoint that records each call as it executes, like zambo.dev/mcp, and the proof is produced by the run itself instead of reconstructed afterward.

Does a checked step prove the agent did good work?

No. It proves the step ran and its saved record is intact. Good is a judgment you make by reading the checked outputs. The record makes that judgment possible; it does not make it for you.