Skip to content

How It Works: Circuit Breaker State Machine

LongGuard borrows the classic Circuit Breaker pattern from distributed systems (like Netflix Hystrix) and reimagines it for LLM cognitive reasoning.

The 4-State Automaton

On every step of an agent's run, LongGuard checks all registered loop detectors and navigates a state machine:

LongGuard Architecture

1. CLOSED (Normal Operation)

  • The agent is executing without detected loops or abnormal token spikes.
  • All detectors run concurrently in sub-milliseconds.
  • Every check returns BreakerDecision(action="continue").

2. REFLECTING (First Loop Detected)

  • A detector has identified a loop pattern with high confidence (e.g. 3 identical searches in a row).
  • Instead of terminating the agent, LongGuard injects a Reflect & Pivot system prompt into the agent context.
  • Returns BreakerDecision(action="reflect", inject_prompt="...").

3. HALF_OPEN (The Recovery Window)

  • The agent has been prompted to pivot.
  • LongGuard monitors the next step:
  • Success: If the agent switches strategies or makes meaningful progress, the breaker resets back to CLOSED.
  • Failure: If the agent ignores the prompt and repeats the bad pattern, the breaker trips to OPEN.

4. OPEN (Graceful Termination)

  • The agent was unable to recover.
  • LongGuard kills the agent cleanly:
  • Saves the entire message history and conversation state.
  • Appends an explanatory AI message to the state so the user isn't left with a broken error.
  • Finalizes the GuardReport with exact timestamps, steps, and kill reasons.

Hard Caps

In addition to loop detection, LongGuard enforces hard safety caps that bypass the state machine: - Token Budget (max_tokens_per_run): Trips immediately if total query tokens exceed the threshold (e.g. 50,000 tokens). - Step Limit (max_steps): Trips immediately if total reasoning steps exceed the threshold (e.g. 25 steps).