As of September 2026, a lead agent in my repo writes a brief for each delivery agent, listing the design decisions already made. In one wave of six runs, one sentence in that brief changed what the agents did.

#Two sentences, then a third

The briefs said: these decisions are made, implement them. If one is wrong, not merely different, stop and say so with the evidence. The first sentence stops an agent from re-arguing a settled design. The second stops silent compliance.

It half worked. One agent found a starvation flaw in the design it was handed. It built the design anyway and described the flaw in its report. It had done what it was told.

So a third sentence went in. If you can prove a decision wrong with a failing test, make the smallest fix and lead your report with it. If the fix changes the scope, stop and report.

#What six runs did

The sentence landed between the wave's first three runs and its last three. The first three corrected nothing. The last three each corrected at least one briefed decision. One run also found that a size limit it had added itself let a read of 20 of 58 products report itself as complete, and took the limit out.

#What this does not show

Six runs in one wave, an accident rather than a trial. The later runs had different tasks, and the same change also edited the delivery agents' own instructions. I have not re-run the same tasks without the sentence. It costs three lines, so it stays in every brief.

These decisions are made. Implement them rather than re-deciding.
If one is WRONG (not merely different from what you would pick), stop and say so in your report with the evidence.
If you can PROVE it wrong, with a test that fails against the briefed design, make the smallest fix and lead your report with what you changed and why.
Do not ship a defect you have proven.
If the fix changes the task's shape or scope, stop and report instead.