Multi-agent "toxic boss"
· Source: original
🎭 Multi-agent "toxic boss": shouting didn't help, the acceptance contract did
The author rebuilt the role of "AI office boss" within a multi-agent system several times. The first hypothesis: the result would be driven by a tough role personality — that very "toxic boss."
A comparison of the instructions revealed a different cause. The improvement came from three elements: accountability for the outcome, specific returns, and a ban on accepting work with unresolved defects. The working formulation of the role:
"the right to say FAIL"
The author calls the mechanism that worked a proper acceptance contract, not shouting. In a multi-agent setup, acceptance is held together by mechanics, not tone. There are no numbers of agents, no model names, no versions, and no metrics in the breakdown — the instruction edits are entirely in the breakdown on Habr