An agent that did not write it

Think of it when an agent says its own change is correct and nobody else has looked.

In everyday life

It is like asking a friend to proofread your letter. You read what you meant to write; your friend reads what is on the page.

One case

An agent adds a renewal limit to a library app: a loan may be renewed at most 3 times. It reviews its own change twice and finds nothing. A second agent on a different model reads the change once and shows, with a failing test, that a 4th renewal still goes through. The author had tested only the first and second renewals.

An agent that did not write it
The ideaCheck it
An agent that did not write it

A second agent that reviews a change it did not write, with a fresh window and often a different model. It looks for the problems the author cannot see.

Check it, Steering

Every case

An author reviews its work with the same assumptions that produced it, so it misses the same things twice. A reviewer started fresh sees the change and the goal, not the author’s reasoning. A different modelThe trained AI system behind an agent. Different models, often from different companies, make different mistakes. brings different blind spots, so a problem that two independent reviewers both name is more likely to be real. Reviewers also report noise and push for extra work, so keep a finding only when it comes with proof, such as a failing test.

The word

Engineers call this

reviewer agentA second agent that reviews a change it did not write, with a fresh window and often a different modelThe trained AI system behind an agent. Different models, often from different companies, make different mistakes.. It looks for the problems the author cannot see.

Have each change reviewed by an agent that did not write it, ideally on a different model. Accept the findings it can prove, and refuse the ones that only grow the change.

Where this idea leads

Skills that encode it

  • interrogate pstackAdversaries try to break the change, and a finding survives only when a run proves it.
  • council tstackThree strong models from three providers review the work, and the lead fights to keep scope limited.
  • sol-loop tstackA different model reviews in a loop with a round budget; make minimal changes and push back on scope creep.
  • show-me-your-work pstackA subagent on a different model family reviews the trail before handback: self-review is not a substitute.