IBM Research and UC Berkeley applied MAST (Multi-Agent System Failure Taxonomy) to ITBench—an enterprise benchmark for SRE, Security, and FinOps automation—to diagnose why LLM-based agents fail in real-world IT workflows. By annotating 310 execution traces across Gemini-3-Flash, Kimi-K2, and GPT-OSS-120B, they found that stronger frontier models fail cleanly with ~2.6 failure modes per trace, while weaker open models cascade with up to 5.3. The most universally fatal failure is FM-3.3 (Incorrect Verification)—agents declaring success without checking ground truth. Kimi-K2 struggles with termination awareness, while GPT-OSS-120B suffers from memory loss and reasoning-action mismatch. Practical fixes include externalizing verification via tool-mediated evidence, implementing finite state machines for termination control, and aggressive context hygiene for models prone to cascading failures. Prompt engineering alone yields only ~15% improvement; architectural interventions like Summarizer Agents or State Machines can yield up to 53%.
