Five failure modes in production agent loops
An observational study of 200 real agent runs — cataloguing the repeating failure patterns that kill multi-step tasks before the finish line.
Multi-step agents fail in predictable ways. The failure modes are not random — they are structural, and they recur regardless of the underlying model.
Over six weeks we logged 200 complete agent runs across three task domains (code generation, research synthesis, document QA). We coded every failure by type: context window overflow, tool hallucination, loop detection failure, premature termination, and instruction drift.
Context window issues account for 38% of failures — far more than tool hallucination (22%) or loop detection (18%). The implication is that memory and context management is a more important engineering lever than prompt quality for long-horizon tasks.
We also found that task complexity, measured by step count, follows a power law: adding one step more than doubles the probability of failure. Architectures that decompose tasks early and validate intermediate outputs dramatically outperform those that commit to a long plan at the start.