The three taps, replaced by one check
Tap one: pick the claim that matters. Not the whole job, the single claim you would be most annoyed to learn was wrong: the payment that went out, the file that was sent, the booking with the right dates. Approval fatigue is real, and checking everything is how people end up checking nothing. One claim, the expensive one.
Tap two: open its receipt. Work routed through a receipt layer carries one receipt per finished step, each at a public URL, each fixing the hash of the saved result at the moment the work ran. If the summary names a payment, there should be a receipt whose record is that payment. If there is no receipt, the claim is a sentence the agent wrote. Sentences are cheap; that is the entire problem the screen was supposed to solve.
Tap three: recompute it. Paste the receipt into the verifier. A match means the saved result has not changed since it was saved: the record you are reading is the record the work left behind. A mismatch, or a missing receipt, means approve nothing yet. Total elapsed time once the habit forms: about a minute, spent exactly where the risk is.
Steering is not verifying, and the screen does only one of them
Approval gates earn their place: an agent that pauses before sending, spending, or deleting is an agent you can delegate to. But a gate answers may it proceed, and the summary on it cannot answer did it do what it says it did, because the describer and the doer are the same system. The industry spent this month arguing about how often agents should nudge you and how much control you should keep. Almost nobody is arguing about proof, which is the half of trust that does not depend on the agent's personality. Build the proof habit now, while the delegation is still small: receipts behind claims, one recomputed spot check per approval, and unverified claims read as unverified.
If your agent's app shows no receipts
Most do not. The consumer personal agents in this month's wave (Claude Duet, Gemini Spark, Meta's Muse Spark, OpenAI's dots, and Microsoft Copilot) are built on review and approval, and none of them documents a portable, independently checkable receipt for a finished job. The practical move is to route the consequential work through a layer that does. The Zambo endpoint adds as a custom connector by URL in the clients that accept one, and every call made through it returns an AER-1 receipt with a public URL. The approval screens keep doing their job, steering. The receipts do the other job, proving, and your yes starts to mean something again.
The honest limit, stated on every page here: a receipt proves the saved result was not changed after it was saved. It does not prove the agent chose well. You still read the work. You just stop having to take its word for what the work was.
Questions, answered
What should I check before approving my AI agent's work?
Check the record, not the summary. The approval screen shows the agent's own account of what it did, written by the system asking for your trust. Ask for the receipts behind the claims: one verifiable receipt per finished step, each with a hash fixed when the work ran. Open one and recompute it at a public verifier like zambo.dev/verify; a match means the saved result has not changed since it was saved. Then read the claims with no receipt behind them as exactly what they are: unverified.
Why is the approval screen not enough on its own?
Because it collapses two different acts into one tap. Steering is deciding what the agent may do next, and approval screens are good at that. Verification is establishing what the agent already did, and a summary cannot do that for itself: the writer and the worker are the same system. When the only evidence for finished work is the worker's description of it, approving is rubber-stamping with extra steps. A receipt separates the acts: steer with the screen, verify with the record.
What if my agent's app has no receipts to show me?
Then treat every summary as a claim, and route the work that matters through a layer that receipts it. Adding the Zambo connector (https://zambo.dev/mcp) to the AI you already use means the calls it makes there come back with an AER-1 receipt each: the call, its time, and the hash of its saved result, at a public URL. Apps differ in what they surface. The consumer personal agents in this month's wave (Claude Duet, Gemini Spark, Meta's Muse Spark, OpenAI's dots, and Microsoft Copilot) are review-and-approval based and do not document portable receipts of their own, so the connector is the route that exists now.
How long does the check actually take?
About a minute for the spot check that matters: pick the claim you would be most annoyed to discover was wrong, open its receipt, and recompute the hash at the verifier. You are not auditing every step; you are establishing that the record exists, that it recomputes, and that claims in the summary map to entries in the record. An agent whose work receipts cleanly earns faster approvals over time. An agent whose summaries outrun its receipts has told you something important, cheaply.