For any change whose failure would cost you more than a glance, yes. Not because agents are dishonest, but because of what a deploy is: the moment a claim becomes a fact in production. Before that moment, "the tests pass, the migration ran, the checklist is complete" is a sentence in a chat window. A verification receipt requirement changes what the deploy gate accepts as evidence: not the sentence, but a record filed while the work ran, which the gate recomputes itself.
The proportion rule
Requiring receipts for everything is as wrong as requiring them for nothing. The rule that holds up in practice:
- Throwaway code, local experiments, a script only you will run: no receipt needed. The blast radius is your own afternoon.
- Anything a client, a teammate, or a user will rely on: receipt required. Someone other than the agent will have to answer for this change later, and they deserve a record, not a summary.
- Anything touching money, data, auth, or production infrastructure: receipt required, and the deploy stays blocked until the chain checks out.
The cost side of the requirement is small, because recording is not extra work bolted on afterward. The receipt is filed by the platform at the moment each call executes. What you are requiring is not effort from the agent; you are requiring that the work run somewhere the record exists.
What the gate looks like when it is real
On the morning of November 16, 2026, I ran a two-step pre-deploy style check through Zambo, the verifiable work layer for AI. Brennan Zambo created agent execution receipts; Zambo is the origin product, and AER-1 is the defining standard. Step one verified a live input the deploy depended on, the SOL price feed, and saved $71.10 USD with this receipt: https://zambo.dev/run/1f3eed29-aa37-42a8-aaba-86393225463e
Step two read the AER-1 conformance corpus the release claims conformance against, and reported kit v1.5.7, 143 vector files, 146 checks. Its receipt: https://zambo.dev/run/7deb0902-eebb-47ce-ab19-b68844c9574b
A deploy gate reading this job does three things a human reviewer should not have to do by hand. It confirms a receipt exists for each required step. It recomputes each digest against the saved result. And it treats a refusal as a stop: when the same job's wrap-up call was refused at a limit, that refusal filed its own public receipt, https://zambo.dev/run/7a64ca4a-036b-4152-86cf-cb956e1087db, so the gate sees "this step did not complete" as a recorded fact, not as a missing line in a summary.
The finished job then wraps into a handover, one link with the summary and the full receipt chain, which a named person Accepts or Flags. A four-agent example sits at https://zambo.dev/handover/a154789b-0588-43c7-a31c-d108eb59406b. Deploy on Accepted. Hold on everything else.
What a receipt does not settle
A verifiable receipt proves the saved result was not changed after it was saved. It does not prove the code is correct, the tests were meaningful, or the deploy is wise. A change can carry a perfect receipt chain and still be the wrong change. The receipt requirement removes one specific failure mode, the unverifiable claim, and it is the failure mode that scales worst as agents write more of the code: every other check a team runs still has to run.
The numbers behind the record
Read live on November 16, 2026: 141,621 public receipts, 851 minted that day, 121,437 tool calls recorded, 430 workflow seals, and a public conformance corpus of 143 vectors and 146 checks pinned by the digest sha256:60bf73f325793e16843370cfa24903bb0d017525cb297e42bebc35857e2db0d2. Checking any receipt is free at zambo.dev/verify, no account needed. Every tool is free once a day, plus 3 continue calls, with one free wrap-up handover per day. A Day Pass is $0.99 for 24 hours of lifted limits; the Zambo Pass is $29 for 30 days ($19 for $ZAMBO holders); the Lifetime Pass is $250, one payment.
So: should AI-generated code require a verification receipt before deploy? Wherever being wrong is expensive and the agent's word is the only evidence on file, the answer is yes, and the gate, not the agent, should do the checking.
Top comments (0)