Model or Harness? An Interaction-Centric Taxonomy for Localizing Agent Failures
The paper argues agent failures should be assigned to the interaction where they start, not just to the final bad outcome.
Its taxonomy maps 41 failure modes onto edges between components such as models, harnesses, users, tools, memory, and environments. Each failure also gets a “fault side,” pointing to whether the fix belongs in post-training, scaffolding, tool integration, environment design, or grading. The authors say this structure is meant to work across coding agents, long-horizon assistants, and multi-agent systems. They report reproducibility checks using reasoning agents as judges, with the strongest frontier model reaching Cohen’s kappa of 0.76 against human labels. Source: HF Daily Papers' note.
Its taxonomy maps 41 failure modes onto edges between components such as models, harnesses, users, tools, memory, and environments. Each failure also gets a “fault side,” pointing to whether the fix belongs in post-training, scaffolding, tool integration, environment design, or grading. The authors say this structure is meant to work across coding agents, long-horizon assistants, and multi-agent systems. They report reproducibility checks using reasoning agents as judges, with the strongest frontier model reaching Cohen’s kappa of 0.76 against human labels. Source: HF Daily Papers' note.
score 4