New Benchmark for Detecting AI Agent Runtime Faults.

Chenkai Zhang, Yiran Li, Yifang Tian, Michalis Bachras, Hans-Arno Jacobsen· August 18, 2026 View original

Key takeaways

  • Reliability in AI agent systems requires diagnosing runtime faults, not just final outcomes.
  • AGENTCHAOSBENCH is a new benchmark for detecting and localizing these faults from telemetry.
  • Current LLMs perform poorly at diagnosing complex agentic system failures from telemetry.
  • Reference-dependent faults, like bypassed guardrails, are particularly challenging to detect.

Who benefits

Software DevelopmentAI/ML EngineeringDevOpsCybersecurityIT Operations

Summary

AGENTCHAOSBENCH is a new benchmark for detecting and localizing runtime faults in LLM-based agentic systems using execution telemetry, revealing that current zero-shot LLM baselines struggle significantly with fault diagnosis, especially for reference-dependent issues like bypassed guardrails.

A new benchmark, AGENTCHAOSBENCH, has been introduced to address the challenge of diagnosing failures in LLM-based agentic systems. Unlike evaluating only final task outcomes, this benchmark focuses on detecting and localizing runtime faults from the system's execution telemetry, which includes tool calls, model calls, guardrails, and inter-agent messages. The benchmark involves injecting ten types of operational faults into five diverse agentic applications, creating a dataset of 275 sanitized traces. Initial evaluations using zero-shot LLM baselines (up to 14B parameters) showed poor performance, with top-1 fault-type accuracy ranging from 13.6-19.2% and joint fault type/location accuracy topping out at 22%. Reference-dependent faults, particularly bypassed guardrails, remained largely unsolved, indicating that current LLMs struggle significantly with diagnosing complex agentic system failures from telemetry alone.

Why it matters

For professionals developing, deploying, or operating AI agent systems, this research highlights a critical gap in current diagnostic capabilities, emphasizing the need for better tools and methods to ensure reliability and troubleshoot failures effectively.

How to implement this in your domain

  1. 1Prioritize the development of robust telemetry and logging for AI agent systems to capture detailed execution traces.
  2. 2Investigate and implement advanced fault detection and localization techniques beyond simple LLM zero-shot inference for agentic systems.
  3. 3Develop specific testing scenarios that simulate various operational faults, including tool failures and guardrail bypasses.
  4. 4Contribute to or utilize benchmarks like AGENTCHAOSBENCH to rigorously evaluate the reliability and diagnostic capabilities of agentic AI.

Original post by Chenkai Zhang, Yiran Li, Yifang Tian, Michalis Bachras, Hans-Arno Jacobsen

"arXiv:2608.14680v1 Announce Type: new Abstract: Reliability in LLM-based agentic systems is a property of the whole execution (its tool calls, model calls, guardrails, and inter-agent messages), not of the final answer alone, yet evaluating only task outcomes reveals little about…"

View on X

Originally posted by Chenkai Zhang, Yiran Li, Yifang Tian, Michalis Bachras, Hans-Arno Jacobsen on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses