Minimum Sufficient Collaboration in Loop Agents

Loop engineering turns agent design into a control problem: systems must discover work, act with tools, verify state, remember traces, and decide whether another cycle is needed. We study adaptive collaboration in loop agents, where the controller chooses between non-collaborative execution, deterministic verification, and role-separated collaboration based on the task contract and workspace evidence. Our contribution is a failure-driven controller for minimum sufficient collaboration: use fast paths for directly observable answers, deterministic gates for locally checkable contracts, and actor/reviewer collaboration only when independent verification can catch omissions that local checks miss. Across small checkable tool and artifact slices, the evidence is mixed but informative: AdaptiveLoop leads a clean Harness-Bench artifact slice and improves τ-bench retail slice, while stronger review/search baselines win on a later independent artifact-audit repeat and deterministic gates alone solve a RealFin-style cache slice. Loop agents should spend collaboration only when it adds independent verification.