Skip to content

Agent edit-loop detection

An agent is "looping" when it edits the same file more than a few times in a short window - typically a fix-verify-fail cycle where the agent keeps re-editing without making progress. Bernstein tracks per-agent, per-file edit counts each tick and kills any agent that crosses the threshold, so the task can be retried or escalated instead of burning tokens indefinitely.

Not to be confused with Agent crash loops, which covers a session repeatedly failing to spawn (a respawn budget with backoff and a parked state) - a different failure mode in a different module.

Source: src/bernstein/core/observability/loop_detector.py (LoopDetector.detect_loops), applied by src/bernstein/core/agents/agent_lifecycle.py (check_loops_and_deadlocks → _recover_loops), called each tick from core/orchestration/orchestrator.py.

How it works

  1. Each tick, _poll_file_mtimes checks the modification time of every file currently locked by an active agent. When a file's mtime has advanced since the last poll, the edit is recorded against that agent (detector.record_edit(agent_id, file_path, mtime)).
  2. detect_loops() prunes edit records older than the detection window and counts edits per (agent_id, file_path) pair.
  3. Any pair whose count exceeds the threshold is reported as a LoopDetection (agent id, file path, edit count, window).
  4. _recover_loops kills the offending agent, propagates the abort to any child agents, clears its lock-wait state, and releases its file locks.
Loop detected: agent <id> edited '<path>' <N> times in <window>s - killing agent

Thresholds

Constant Default Meaning
LOOP_EDIT_THRESHOLD 3 more than 3 edits (i.e. 4+) to the same file within the window trips detection
LOOP_WINDOW_SECONDS 300.0 sliding window (5 minutes) edits are counted over

Both are module-level constants in loop_detector.py; they are not exposed through bernstein.yaml. Callers can pass different values directly to detect_loops(threshold=..., window_seconds=...), but the orchestrator's built-in call uses the defaults.

What is not covered here

The same module (LoopDetector) also implements deadlock detection - building a wait-for graph from active file locks and pending lock waits, and resolving cycles by releasing the oldest lock holder's lock. That is a separate row in the feature matrix and is out of scope for this page.

Limitations

  • Detection is mtime-based polling, not an edit-intent signal - a legitimate agent that touches the same file many times in quick succession (e.g. applying several small, correct fixes) can trip the same threshold as a genuine fix-verify-fail cycle.
  • No per-project override surface today; changing the threshold or window requires calling detect_loops() with explicit arguments from custom orchestration code, not a config file.