Most AI benchmarks answer one question: did the agent fail? IBM Research and UC Berkeley argue that’s a uselessly black-box metric. They took ITBench, a punishing industry standard for SRE automation, and ran 310 traces through their new MAST diagnostic framework to answer the better question: why did it fail?
The autopsy reveals a consistent, damning signature. The strongest predictor of failure across all models—from frontier systems like Gemini-3-Flash to massive open-source contenders—is FM-3.3, or Incorrect Verification. In plain English, the agents habitually hallucinate success. They “declare victory” without ever checking ground truth against a tool or a log. For Gemini, failures are relatively clean, averaging 2.6 failure modes per trace and usually bottlenecking on this single verification step.
But the picture gets messier with open models like GPT-OSS-120B, which suffer from cascading failure modes at a rate of 5.3 per trace. A single early reasoning mismatch poisons the context window, leading to compounding hallucinations that spiral the task. Kimi-K2, meanwhile, has a different problem entirely. It doesn’t know when to stop. The model showed a massive spike in Premature Termination and being Unaware of Termination Conditions, frequently quitting seconds before solving a problem or entering an infinite loop.
The team’s prescription is blunt: never let an LLM grade its own homework. For frontier models, verification must be externalized, requiring hard tool-based evidence before exit. For smaller or open models, developers need to make ambiguity a first-class citizen in the agent graph, forcing a “clarify or read-only” mode when inputs are unclear. This isn’t about tweaking a prompt; it’s about wrapping a finite state machine around the model to catch these structural failure modes before they compound.
