Signals triages agent traces without LLM judges

This is the kind of infra paper that matters if you are already drowning in agent logs and eval traces.

• –The core idea is pragmatic: move triage upstream with cheap signals instead of paying for human or LLM review on every run.
• –The paper’s value is less about a new model and more about a better sampling layer for agent observability and post-deployment improvement.
• –The no-GPU, no-online-behavior-change constraint makes it easier to slot into production systems.
• –The main limitation is obvious: signal heuristics will miss some subtle failures, so this is best viewed as a prioritization layer, not a replacement for deeper review.