Detecting and Breaking Infinite Loops

One of the most common failure modes of autonomous agents is getting stuck in repetitive loops: repeatedly searching for the same failing query or calling the same failing tool. Robust production systems must have hard programmatic loop breakers.

The Loop Breaker Pattern

• Max Iterations Cap: Always enforce a hard ceiling (e.g. 10 or 15 iterations maximum) regardless of what the LLM desires.
• Action Fingerprinting: Hash the combination of `tool_name + JSON.stringify(args)`. If the exact same hash occurs 3 times consecutively, intervene immediately.
• Nudge Prompt: When a loop is detected, inject a system message: 'Warning: You have repeated this exact action without new progress. Choose a different approach or report your blocker to the user.'

const historyHashes = new Set<string>();
for (let step = 0; step < MAX_STEPS; step++) {
  const action = await getNextAction();
  const hash = `${action.tool}:${JSON.stringify(action.args)}`;
  if (historyHashes.has(hash)) {
    injectSystemMessage('Loop detected! Do not repeat previous actions.');
  }
  historyHashes.add(hash);
}