Skip to content

Circuit Breakers for Agent Loops

Circuit breakers stop agent loops when progress stalls — repeated errors, escalating costs, context exhaustion, or circular behavior signal a halt rather than continuation.

Also known as: Loop Detection & Stopping. For the complementary technique of detecting repetitive edits within a session, see Loop Detection.

The Loop Problem

Agents in open-ended loops consume resources without progress — applying the same incorrect fix, retrying a flaky test 20 times. Without stopping conditions, loops run until the context window fills or the session is killed, leaving degraded context and unusable partial output.

Stopping Signals

Five signals warrant a circuit break:

1. Iteration limit reached. The agent has taken N steps without completing the task. N varies by task type. Claude Code sub-agents support a maxTurns field (docs) that enforces this at the runtime level.

2. Repeated failure. The same tool call fails repeatedly with the same error. A 429 three times in a row will continue to 429. A test failing for a logic error will continue to fail.

3. Repetition detected. The agent is doing the same thing it already did — fetching the same URL, reading the same file, attempting the same fix. Repetition without new information is a stuck loop.

4. Context budget exceeded. The context window is approaching the "dumb zone" where output quality degrades. Chroma's Context Rot study tested 18 frontier models including GPT-4.1, Claude Opus 4, and Gemini 2.5, and found every model degrades non-uniformly as input grows, with onset depending on task similarity and distractor density. Trip the breaker when recall or coherence drops on your task, not at a fixed token count.

5. Cost threshold exceeded. The task has consumed more than the expected budget — cost overrun often correlates with loops.

Graceful Degradation

When a circuit breaker trips, the agent should:

  1. Stop executing new actions
  2. Return partial results completed before the break
  3. Explain what triggered the stop and what remains incomplete
  4. Escalate to a human if the pipeline has a human gate

Partial results are more useful than nothing. Return what you have; do not discard completed work.

Configuration

Signal Configuration Enforcement
Iteration limit maxTurns in agent frontmatter Runtime
Cost threshold Session budget settings Runtime
Error rate Agent instruction + hook Instruction / hook
Repetition Agent instruction + hook Instruction / hook
Context usage Agent instruction Instruction

Runtime enforcement (maxTurns, cost budgets) cannot be overridden by the model. Instruction-level enforcement depends on the model obeying instructions — less reliable for safety-critical stops. Hooks offer a middle ground: deterministic scripts that monitor and trigger a stop.

Circuit breakers are the enforcement mechanism for context health — without them, context management guidelines are advisory rather than operational limits.

Example

The following Claude Code sub-agent definition combines a runtime-enforced maxTurns limit with an instruction-level check for repeated failure. Both signals are present so that the agent stops whether the loop is caused by hitting the turn ceiling or by a recurring error pattern.

# .claude/agents/research-agent.md frontmatter
---
name: research-agent
description: Fetches and summarises web sources for a given topic
tools:
  - WebFetch
  - Read
  - Write
maxTurns: 20
---

# Research Agent System Prompt

You are a research agent. Fetch and summarise up to 5 sources for the given topic.

## Circuit-breaker rules

1. **Iteration limit** — enforced by `maxTurns: 20` above; you will be stopped automatically.
2. **Repeated failure** — if the same URL returns an error three times in a row, skip it,
   note it as unreachable, and move to the next source. Do not retry indefinitely.
3. **Repetition detection** — if you find yourself fetching a URL you have already fetched
   this session, stop and return what you have collected so far.
4. **Partial results** — when you stop for any reason before completing all 5 sources,
   return the summaries you have already written plus a short note explaining what
   triggered the stop and which sources were not completed.

The maxTurns: 20 field is enforced at the Claude Code runtime level and cannot be overridden by model reasoning. The instruction-level checks handle error-rate and repetition signals, which the runtime does not detect automatically.

When This Backfires

Circuit breakers are failure-mode detectors, not correctness guarantees — set too aggressively, they become the failure mode. A reasonable practitioner would push back in at least three situations:

  • Iteration limits trip on legitimate work. Setting maxTurns low enough to catch pathological loops also cuts off legitimate multi-step refactors or research tasks — several production frameworks have open issues where agents halt mid-task on "stopped due to max iterations" even when making forward progress (openai-agents-python#844, langflow#10607). Raise the ceiling for task classes that legitimately need 50+ turns.
  • Repetition detection flags valid re-reads. Re-reading the same file after an edit, or refetching a URL after a 429 backoff, are normal behaviors, not stuck loops. Naive "did we already fetch this?" heuristics fire on both.
  • Cost thresholds penalize exploration. Exploratory research agents legitimately consume variable budgets. A hard cost cap trips on successful discovery runs as readily as on loops — the signal is cost without progress, not cost alone.
  • Instruction-level stops are model-dependent. Signals 2, 3, and 5 rely on the model reading its own circuit-breaker rules and obeying them. If the model ignores the instruction mid-reasoning, the stop never fires. For safety-critical stops, prefer runtime enforcement (maxTurns, hooks) over instructions.

The steelman: if your agent population already fails gracefully on its own — returns partial results, detects its own thrash — adding another stopping layer mostly creates false positives. Instrument first; add breakers where instrumentation shows real loops, not prophylactically.

Key Takeaways

  • Five stopping signals: iteration limit, repeated failure, repetition, context budget, cost threshold
  • maxTurns provides runtime-enforced iteration limits; instruction-based checks can be overridden by the model
  • Graceful degradation: return partial results + failure explanation, never discard completed work
Adopted — practitioner-sourced, corroborated Last reviewed: 2026-06-13
Feedback