DEV Community

Cover image for My agents could edit the checks that judged their work, and the checks still passed
Justin Bartak for Orbyt Labs

Posted on Originally published at orbytlabs.ai AI-assisted

My agents could edit the checks that judged their work, and the checks still passed

An outside review found that my agents could write to the same repository their checks read, so a worker could change its own judge.

  • The fix I documented forbids the reviewer from editing files or even running the suite, because some scripts that look observational actually write committed artifacts. It records the working tree before and after.
  • I haven't shown that the gap is closed. No worker has tried to rewrite a governing check under a separately enforced boundary and been refused, so that proof is still missing.
  • The 664K source lines and 37,623 checks from October 2, 2026 describe an internal build with no paying customers. Size is not demand, and 100x is a claim about scope, not a measured multiplier.

Read the full article: The Blueprint for the 100x Developer

This summary was written with AI assistance (Claude) from the full article at https://www.orbytlabs.ai/blog/blueprint-100x-developer.

Top comments (0)