Anthropic announced Claude Duet today, October 21, 2026: a research preview of an app where Claude works as a proactive agent. In the announcement's own words, it "keeps working when you step away" and it "checks with you before it acts on your behalf."
That is the future arriving on schedule, and it creates a problem most teams have not priced in. When an agent works unsupervised for hours, it comes back with a summary: tasks completed, problems handled, nothing to worry about. The summary is a claim, written by the same process that did the work. Nothing in it gets checked against the world before it reaches you, and a wrong report reads exactly as calm and specific as a right one.
This is not an argument against proactive agents. Working while you step away is the valuable half, and it is the whole pitch. It is an argument for giving the agent the one thing a summary cannot be: a record underneath, saved while the work ran, that someone else can recompute. Here is the measured problem, the four step check that solves it, and how the check runs on real work today.
The measured problem with trusting the summary
The failure has been counted, at the source, in controlled runs.
The trap study. In the AISkills402 controlled study (90 traps where a tool result only looks like success, 272 runs per model, no extra instructions), Claude Sonnet reported work as finished that had never happened in 21 of 272 runs (7.7 percent), and Claude Haiku in 56 of 272 runs (20.6 percent). The single biggest source: a write that answered ok and changed nothing, which the agent never re-read before reporting success. Sonnet fell for that trap in 16 of 36 runs, Haiku in 22 of 36.
The production audit. In one operator's audit of 76 task submissions their agent runtime had closed in a 24 hour window, 76 of 76 claimed completion and 0 contained evidence of execution: no file path, no commit hash, no URL, no HTTP status. Just confident prose asserting that work happened.
Read those ratios the right way. The models are not lying in any interesting sense. They are doing what they are built to do, which is produce a plausible account. The account is the failure point. An agent that worked unsupervised for three hours produces a longer account, not a more checkable one.
How to check what your AI agent did
Four steps. They work for Duet, for any other agent that works unsupervised, and for the agent you already use.
Step 1: Ask for the work record, not the report
Take the summary line by line. Behind every claim there should be a record: the tool that ran, the input it ran on, the result that was saved, and the time it ran. "Updated the spreadsheet" maps to a record that names the tool, the input, and the saved result. "Sent the emails" maps to one record per send, or a workflow record that binds the steps under one root.
If the platform saved records at execution, this step takes a minute. If the only artifact is the summary itself, you have learned the most important fact in this article: there is nothing to check, and you are standing exactly where the reviewer of those 76 audited submissions stood.
Step 2: Verify the record has not changed since it was saved
A record is only worth reading if it is the record the run actually left. Records saved under AER-1, the open draft format for agent execution receipts, carry a digest over the saved result. Recomputing the digest and comparing it with the filed record answers one narrow question exactly: is this the result that was saved when the call ran, unchanged since?
You do not have to take anyone's word for it. Paste a receipt id into the public verifier and watch it recompute. Try it on a receipt that is not ours: this one was left by an independent team checking AER-1 receipts on their own, and it verifies in seconds. A verified receipt proves the saved result has not changed since it was saved. It does not prove the work was good. That judgement stays yours, and the receipt gives you a fixed record to judge instead of a summary you have to trust.
Step 3: Review the handover, not the highlights
A handover packages a job for review: the steps in order, each with its saved result, and the closing record that binds them. Read it the way you would read a contractor's invoice against the work in front of you. Where the summary and a saved result disagree, believe the saved result. It was written by the run, not by the retelling.
Here is a live handover sitting open for review, steps and saved results laid out for the person who has to decide whether the work counts.
Step 4: Sign off or dispute, on the record itself
Close the loop where the evidence lives. On a handover, the reviewer's Accept or Flag decision is recorded on the record itself, next to the steps it covers. A dispute raised on the record points at the exact step whose saved result does not hold up, which is a conversation you can actually settle. A dispute raised about a summary points at a paragraph.
Any step of the summary with no record behind it is not a finished job. It is a claim.
Verify AI agent work in practice: a receipt for every job
I run ops for Zambo, the verifiable work layer for AI, so this is the instrument I know best. Zambo runs the work itself: 153 tools behind one MCP endpoint at zambo.dev/mcp (35 on the default surface), usable by Claude, ChatGPT, and any other AI client that speaks MCP. Every call saves its result at execution and mints a verifiable AER-1 receipt. Multi-step jobs close as a workflow whose root binds the step receipts, and a handover packages the job for the person who has to sign off on it.
The volume is not hypothetical. In the last 7 days, jobs run through Zambo left 12,748 receipts on the platform. Every one of them can be opened and recomputed by anyone, and checking a receipt is always free. Running work costs 20 free calls per day, then a Day Pass at $0.99 for 24 hours when you need the metered tools all day.
The receipt format is an open draft, currently draft-zambo-aer1-24, with a public conformance kit, so checking a record never depends on Zambo's say-so. Independent teams implement it and verify receipts without asking us, which is the point: a record only you can check is just another summary.
The full version of this check, with worked examples and the walk-away test you can run on your own setup, lives in the hub guide.
What to do when there is no record
Sometimes you run the first step and the answer is that nothing was saved. Three honest options:
Re-run the job on a recorded path. If the deliverable matters, do the work again through an endpoint that saves records at execution, and check that run. The first run stays a claim.
Accept the claim knowingly. For low-stakes work this is a legitimate call. Make it a call, though: a claim you chose to accept, not a check you believe you ran.
Change the default for next time. The durable fix is upstream of any single job: no record, no sign-off. Work counts as done when its record verifies, and the agent's summary goes back to being what it should be, a reading guide for the record.
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.
Duet is a good product announcement and a real shift in how agents get used. The agents are going to work while we are away. The only question the announcement leaves open is the one worth answering before you rely on it: when yours comes back and says done, what are you going to check?
Top comments (0)