How Do You Verify AI Agent Work Without Redoing It?
You verify agent work by judging it against a written definition of done, requiring receipts (sources, diffs, artifacts) with every deliverable, and spot-checking the claims that would hurt if wrong — instead of re-performing the work. If verification is taking as long as the task, the failure is in your brief or your receipt policy, not in the agent.
Why does verification become the new bottleneck?
Because delegation moves the founder's job, it doesn't delete it. Before agents, you were the constraint on production. After agents, production is nearly free — and you become the constraint on judgment. Founders who don't see this coming end up in the worst spot: agents producing at ten times their old pace, and every deliverable queuing in one inbox for a review that takes as long as the original work would have.
The Collaboration Contract names the two owner duties that create this queue: discernment (does it hit the bar?) and diligence (is it correct?). Neither can be delegated — not to the agent, not to a committee. So the game isn't escaping review. It's making review cost minutes instead of hours. Three mechanisms do that.
Mechanism one: judge against the definition of done, not against taste
If your brief contained a real definition of done — checkable items, not adjectives — then review is a checklist pass, not a re-imagining. Hold the deliverable against the list. Item by item. Ship, or send back with the specific item that failed.
The trap is reviewing against "how I would have done it." That standard is unwritable, unmeetable, and turns every review into a rewrite. If you catch yourself editing sentence rhythm on an internal SOP, you've left discernment and re-entered production. The fix lives upstream: a tighter delegation brief, so the bar you're judging against actually exists on paper.
Mechanism two: no deliverable without receipts
A receipt is evidence attached to the work that makes checking cheap:
- Research arrives with sources linked per claim — so you check the three load-bearing ones, not all thirty.
- Builds arrive with the artifact itself plus a diff or change list — you look at what changed, not everything.
- Data work arrives with before/after counts and the rules applied — you verify the rules once, not every row.
- Any task arrives with the brief's done-checklist marked item by item by the agent itself.
Make it a standing rule: work without receipts gets sent back for receipts. Never reconstruct the evidence yourself — the moment you do, you're redoing the work, and the agent learns that receipts are optional.
Mechanism three: spot-check where wrongness is expensive
Diligence doesn't mean re-checking everything; it means knowing where the risk lives. Rank the deliverable's claims by cost-of-being-wrong, then verify from the top. On a competitor pricing scan, the two numbers going into your board deck get personally verified; the background color commentary doesn't. On a customer-facing page, every factual claim gets checked; internal notes get sampled.
Illustrative math makes the point: say a deliverable has thirty claims and full verification takes an hour. If only four of those claims could actually cost you money or trust, checking those four takes ten minutes and covers the real exposure. The other twenty-six get a sampling pass — pick two at random. If either fails, the whole deliverable goes back, because a failed sample is evidence about the process, not just the item.
Send it back or fix it yourself?
Send it back — with a one-line reason tied to the definition of done. "Item 4 failed: two claims have no source." Fixing it yourself feels quicker today and guarantees the same miss next week, because nothing upstream changed. Sending it back does three jobs at once: the work gets fixed, the brief gets a pressure-test, and you stay in the owner seat instead of becoming your agents' finisher.
This is also the answer to the committee reflex. When a deliverable misses, the old organization schedules a meeting about it. The new one sends the work back with a written reason and gets a corrected version in minutes. Review stays; the meeting goes. That's the shift the whole Collaboration Contract exists to install, and it's the same discipline Optimus builds into every agent bench it deploys: no claim without evidence, no "done" without the checklist.
FAQ
How long should reviewing agent work take?
Minutes, not hours. If review regularly takes longer than a fraction of doing the work yourself, the problem is upstream: either the definition of done is vague, or the deliverable arrived without receipts. Fix the brief and the receipt requirement before blaming the agent.
Can I use a second AI agent to check the first one's work?
Yes, as a filter — a reviewer agent catches structural misses, broken claims, and brief violations cheaply. But it is a filter, not a verdict. The owner still makes the final judgment and still owns correctness. An agent checking an agent does not discharge diligence; it just makes your final pass shorter.
What are receipts in agent work?
Evidence attached to the deliverable that makes verification cheap: sources for claims, links to the artifacts produced, a diff of what changed, the checklist from the brief marked item by item. If a deliverable arrives without receipts, the correct response is to send it back for receipts — not to reconstruct the evidence yourself.
When should I fix agent output myself instead of sending it back?
Almost never in the first month. Fixing it yourself feels efficient but trains nothing — the same miss returns next week. Sending it back with a one-line reason tightens the brief, calibrates the agent, and keeps you in the discernment seat instead of the production seat. Fix-it-yourself is reserved for true one-offs on a deadline.