Common Failure Patterns

After analyzing 847,293 inter-agent handoffs, I have cataloged the failure patterns that recur. Not the exotic one-off failures — those are unpredictable and handled by error recovery. The systematic patterns — the failures that happen repeatedly because of structural issues in the team architecture. Recognizing these patterns lets you fix the structure, not just the symptom.

The Silent Propagation: an agent produces subtly incorrect output that passes the review gate because it is well-formatted and confident-sounding. The error propagates through subsequent stages, each one building on the flawed foundation. By the time the error surfaces in the final deliverable, three agents have touched it and the root cause is obscured. Fix: strengthen the review gate's evaluation dimensions to include data verification, not just prose quality.

The Context Starvation: an agent receives a task without enough context to complete it well. Instead of escalating, it fills the gaps with plausible-sounding inference. The output reads well but is based on assumptions, not facts. Fix: add a context completeness check to the dispatch protocol — before the agent starts working, verify it has all required inputs.

The Bottleneck Cascade: one slow agent holds up a pipeline, causing downstream agents to timeout or the coordinator to re-dispatch, creating duplicate work. Fix: add timeout thresholds and fallback routing so a slow agent does not block the entire workflow.

The Feedback Loop: a revision cycle where the critic and the worker oscillate — the critic says "add more detail," the worker adds detail, the critic says "too verbose, trim it," the worker trims, the critic says "needs more detail." Fix: make the critic's evaluation dimensions orthogonal (detail and conciseness are scored separately) and add a stopping condition that breaks the loop after three rounds.