Failure Modes and Recovery
What goes wrong with multi-agent systems and what to do about it.
Three failure modes account for almost all multi-agent pain. Doom loops: an agent repeats the same failing step, burning the window without progress — the fix is noticing early (watch the task list stall) and stopping to re-brief, because iteration 12 of a failing approach is never better than a sharper brief. Collision: two parallel agents touch the same files — the fix is structural, disjoint worktrees and task splits, not hope. Context poisoning: a wrong 'fact' enters the run early (a stale doc, a misread error) and every downstream decision inherits it — the fix is auditing the artifact trail for the first wrong belief, then correcting the source.
All three have the same meta-lesson: build the guardrail into how you brief and partition work, rather than reacting run by run. Defensive design beats incident response.