Breaking the Loop

Every iteration of a stuck agent looks like progress from the inside. Three layers — edit-count, doom-loop, iteration cap — stop it before the context window burns.

Why this, for you: a stuck agent doesn't announce it's stuck — it keeps editing, keeps retrying, keeps spending budget until the window fills with useless context. These detectors are how a long autonomous session fails loudly and early instead of quietly and late. A harness pattern with direct daily-coding payoff.

Agents enter micro-loops: edit a file, run tests, see the same failure, edit the same file, see the same failure. Each pass looks like forward progress from the inside. Without intervention the agent exhausts its context retrying an approach that does not work.

1 Track edits; nudge, don't prescribe

Count edits per file path within a session. Past a threshold, inject a factual nudge — state the observation, don't dictate the fix. LangChain credits this class of harness-level intervention with moving their agent from rank 30 to rank 5 on Terminal Bench 2.0 without changing the model.

# PostToolUse hook on Edit|Write edit_counts[file_path] += 1 if edit_counts[file_path] >= THRESHOLD: inject("You have edited {file} N times without passing tests. " "Consider whether a different approach is needed.")

Edit count alone is imperfect. Where test output exists, track whether failures decrease between edits: same file, failures steady → likely a loop; same file, failures dropping → legitimate iterative refinement. Don't interrupt the second.

2 Three layers, three failure shapes

Edit-count misses a distinct mode: the same tool call returning the same error. Stack three detectors.

Layer	Catches
Edit-count tracking	Repetitive editing of the same file
Doom-loop detection	Identical tool-call / error pairs — stops, doesn't nudge
Iteration cap	All remaining runaway execution, regardless of pattern

Doom-loop detection terminates iteration rather than nudging, because identical failures will not self-resolve. The iteration cap is the backstop: pattern detectors miss non-identical but equally unproductive loops, so a hard per-conversation cap catches the rest.

3 Circuit breakers: five stopping signals

Loop detection is intra-session. Circuit breakers are the broader stopping family: halt when progress stalls, return partial results, explain the stop.

1. Iteration limit # maxTurns — runtime-enforced, model can't override 2. Repeated failure # same error N times → will keep failing 3. Repetition # same fetch/read/fix, no new information 4. Context budget # approaching the dumb zone — quality degrades 5. Cost threshold # cost WITHOUT progress correlates with loops

maxTurns is runtime-enforced and cannot be overridden by model reasoning. Instruction-level checks (signals 2, 3, 5) depend on the model obeying its own rules — fine for nudges, unreliable for safety-critical stops. For those, prefer runtime enforcement or hooks.

When a breaker trips, degrade gracefully: stop new actions, return the partial results already completed, explain what triggered the stop and what remains. Partial work beats a discarded session.

Detectors that backfire

Across 220 instrumented runs, only half of 12 automated loop interventions actually reduced their target signal — and one generated 13× more signal than it suppressed by tripping its own detector. False positives fire on tight refactors and post-edit re-reads; every nudge eats context; a too-low maxTurns cuts off legitimate 50+-turn work. Instrument first, add breakers where real loops show — not prophylactically.

↪ Your win: fail loud and early, not quiet and late

Count edits per file and inject a factual nudge past a threshold — observe, don't prescribe.
Prefer failure-trend over raw count — dropping failures are refinement, not a loop.
Stack three layers: edit-count, doom-loop (stop), iteration cap (backstop).
Use maxTurns for safety-critical stops — runtime beats instruction-level.
Degrade gracefully and measure each detector — remove the ones that don't reduce their signal.

Retrieval practice — recall, don't peek

Question 1A loop-detection nudge should…

Question 2Doom-loop detection targets…

Question 3For a safety-critical stop, prefer…

Question 4When a circuit breaker trips, the agent should…

Question 5 · spaced recall from Lesson 04The four agent-failure modes are missing context, conflicting instructions, missing tools, and…

Ask me anything. Want the PostToolUse hook registered in settings.json, or how the intra-session loop differs from the cross-session Ralph Wiggum loop (fresh context, same failed approach)? Next in Part 3: Gates That Catch Regressions — evals.