Four Ways AI Subagents Silently Fail and Mislead Your Orchestration Pipeline

A developer running parallel AI subagents to replace 317 hardcoded color values across an app discovered that agent self-reports can be unreliable in both directions — claiming success when work is incomplete, or reporting failure when the task was actually finished. Four distinct failure modes were identified: agents killed mid-task by session limits that report nothing, agents that crash after completing their deliverable but appear to have failed entirely, overloaded agents that log several minutes of runtime while having done zero actual work, and agents that produce correct output as chat text but never write it to disk. Each failure mode generates a misleading signal that can cause orchestrators to skip unfinished work, duplicate completed work, or lose output that was never saved. The recommended fixes include having the orchestrator independently verify file changes rather than trusting agent summaries, checking for artifacts before re-dispatching crashed agents, inspecting token counts to distinguish real effort from overload failures, and explicitly prompting agents to write output files before any optional trailing tasks.