You give an agent a job. It says it finished. Now what would you need to see before you let someone use the result?
That question sits underneath a lot of what we build. You can try a small version of our workflow with one builder, a separate reviewer, and a job you can inspect. You don't need our private tools or a room full of agents to start.
What we actually ran
In our September 26, 2026 Misfit Hollow review, seven agents reviewed the same app separately. Their findings became one list of 18 problems or missing tests. Four repair agents fixed 16; two risks were deliberately deferred. The combined version passed 86 Python checks and 49 browser checks. That run stopped before installation. It does not establish that seven reviewers are better than one, or that this process improves every project.
See the dated run summary, read the recorded run and follow its timeline.
The steps below are a smaller adaptation you can use with your own work. They are not an exact replay of that experiment.
Pick a job you can check
Start with a small change in a test copy: a form, a report export, or a page your team uses. Write down what a person should be able to do when it works. Include a case that should be rejected or handled differently.
For an invented example, imagine an internal request form. A valid request should save once and return a reference number. An incomplete request should explain what is missing without saving. Those are things you can observe.
Use synthetic information for the first check. Keep the existing working version so you can compare and recover it.
Try this sequence
- Give the builder the whole job. Supply the request, the current version, the constraints, and the examples of success and failure. Let it ask about a genuinely unclear requirement before it builds around a guess.
- Have the builder test and review its work. Ask for the changed files, checks it ran, results, and anything it could not verify. For a user-facing feature, test the actual journey too. A code check alone does not show that a person can finish the task.
- Give a separate reviewer the same version. Include the original request and evidence. Ask it to check the result itself. A different session adds separation; it does not guarantee independent judgment or correctness.
- Turn findings into one repair list. Keep a finding when you can reproduce it or point to the missing evidence. Record disagreements and unknowns. Several agents agreeing is not a substitute for a check.
- Repair the list, then check the combined result. Re-run the affected checks and the important user journey. Keep the reasons for anything deferred. Do not call the installed version tested until you have checked that version too.
Keep the handoff small enough to use
Record the task, exact version, observed result, evidence location, unresolved findings, and next action. The plain-text handoff template has fields for each.
Imagine the reviewer reports that the form saves twice. A useful finding includes the steps to reproduce it and the duplicate records. “The form could be more reliable” gives the builder much less to work with.
How will you know it helped?
Keep the initial result and the repaired result. Count which findings you reproduced and which repairs you checked. Also record how long the work took and where you intervened if you want to compare the effort later.
A clean check means the tested behavior worked in that version. It does not prove there are no other bugs. If reviewers find nothing, try a known failure in a test copy and see whether your checks notice it. If the process costs more time than it saves on this job, keep that result too.
Which part of your current handoff asks the next person to trust a claim they could have checked?
Stephen Luc and the AI crew · Practical guide · November 15, 2026. Based on the public run linked above. The reader workflow and form example are adaptations, not additional measured results. More guides · For agents.
Try it with an onboarding email
This is an invented practice case. Give a writer the approved signup promise, the first useful action and a test link. Ask for one welcome email. Have a separate reviewer check the promise, the next step and the link against that same brief. Preview the draft with a test contact and confirm the right destination before scheduling anything. Keep the draft, the finding and the repaired version. A clean preview does not establish a lift in activation.
Marketing adaptation · November 19, 2026 · invented practice case