When Agents Fail: Model or Harness--Who's to Blame?

arXiv:2610.32483v1 Announce Type: new Abstract: LLM agents combine models with execution harnesses, so agentic benchmark performance reflects their joint behavior. We argue that harnesses should therefore be evaluated through their interactions with underlying