What is a circuit breaker for AI agents?
A circuit breaker stops an AI agent from retrying a tool call that cannot succeed, ending retry loops before they burn tokens and time.
Updated
A circuit breaker for AI agents stops an agent from retrying a tool call that cannot succeed, ending the loop early so the agent, or its code, can hand off instead of burning tokens and time.
The retry-loop problem
Agents told to "keep trying" will retry a failing call with small variations until they hit a step or budget limit. When the failure is permanent (a missing permission, a closed account, an unsupported operation), every retry is waste, and sometimes harmful.
How an agent circuit breaker works
- Count repeats: the same failure several times in one run trips the breaker.
- Use history: a failure known to be unrecoverable trips it on the first occurrence.
- Return what is known: for a failure with a known fix, return the fix instead of a bare error.
- Raise clearly: a distinct error lets the calling code escalate, hand off to a person or report.
In Retrack
memory.protect(tools) wraps LangChain tools with a breaker driven by the agent's failure signatures. See Known fixes and the circuit breaker, and how to make AI agents cost-efficient for the other places agent spend leaks.