Intel / article

AgentChaos: Chaos Engineering for Agent Systems via Programmatic Fault Injection

Source: arxiv.org
Author: Gou Tan, Zhensu Sun, Jieke Shi, Ting Zhang, Zilong He, Qingfu Wu, Shuai Liang, Weifeng Sun, Junda He, Pengfei Chen, Chuanfu Zhang, Lwin Khin Shar, David Lo
Date: Published Aug 10, 2026

Context

AgentChaos tests what happens when the model service gives an a broken response. It intercepts HTTP API traffic and injects failures such as missing content or corrupted tool-call fields, without requiring changes to the agent's source code. It also checks whether the intended fault actually occurred, so unaffected runs do not make a system appear more resilient. Research abstract

Across 65 fault configurations, the authors report declines of up to 50 percentage points in first-attempt task success. System rankings remained consistent across different underlying models, suggesting that the surrounding agent implementation contributes substantially to resilience. The method makes failure handling something that can be tested directly alongside normal task performance. Reported findings

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source arxiv.org