AgentryWorks · how the work runs
“No agent approves its own work” is a rule of how changes get reviewed here — not a claim that software cannot be attacked. This page states what it does, and what it does not.
On the ordinary pull-request path, each change is reviewed by two separate model gates before it can land. Neither steps aside because the other already looked, and neither can approve a change to its own review surface — that case fails rather than passes.
What gets built, what does not, and how completion is judged are settled up front. A change to that scope is re-agreed rather than absorbed quietly.
The choices that shape a system — and the ones that were rejected, with the reason — are written down, so a later question has an answer that does not depend on anyone’s memory.
Administrative override, direct write access, and a forged status check all sit outside what a review gate can see. The rule constrains the ordinary path that changes travel, which is where mistakes actually happen — not a determined attacker.
This describes how work is run here. It is not an assertion about a client’s infrastructure, and nothing on this site claims otherwise.
Two reviews catch more than one. Neither catches everything, which is why changes that are difficult to reverse get deliberate extra checking before they run.
Questions: studio@agentryworks.com · Back to the front page