AgentChaos framework stress-tests LLM agents, exposing 50-point pass rate drops

Agent systems depend on LLM APIs for nearly every response, yet those APIs can fail in messy ways: server errors, truncated outputs, or corrupted tool-call fields. These faults cascade through downstream agents and derail tasks, but existing testing methods are offline, invasive, or too coarse to mimic realistic failures. AgentChaos, developed by Gou Tan and 12 collaborators and accepted at ASE 2026, takes a different approach: it intercepts and modifies LLM API responses at runtime, at the shared HTTP layer that every agent system uses. That makes it non-intrusive—no source code changes—and lets it target specific response fields with three fault types: crash, omission, and value faults, applied to both content and tool calls. The framework also verifies whether injected faults actually triggered, filtering out untriggered tasks so the measured impact isn't underestimated.

The evaluations were brutal. Across 65 fault configurations, multiple agent systems, and backbone models, every system degraded under fault injection. Pass@1—the rate at which agents complete tasks correctly on the first attempt—fell by up to 50 percentage points. Ranking of systems stayed consistent across different models, which suggests robustness is driven by architectural and prompt-level choices, not raw model capability. Another finding: current fault diagnosis methods are weak, achieving under 53% accuracy on fault type and under 56% on fault step. For engineering teams, AgentChaos offers a practical, repeatable way to proactively test how their agents fail, identify weak links, and improve reliability before real users hit broken APIs. The framework's programmatic fault injection turns an unpredictable problem into a testable one.