Burn-Rate Circuit Breaker for Agent Fleets

MohamedKadri_ · reddit · 2026-07-13

The author shares a burn-rate circuit breaker designed for LLM agent fleets: when an agent's failure rate consistently worsens relative to its own historical baseline, the system automatically demotes it to propose-only mode, generating suggestions without directly executing actions.

Background

While single-action guardrails can block immediate bad calls, they fail to catch agents that are "slowly degrading"—where each output passes inspection, but the overall failure rate is quietly rising.

Core Concept

Inspired by telecom NOC and Google SRE burn-rate alerts:

Track a trailing baseline of failure rates for each agent.
Compare the current window against the agent's own historical performance.
Trigger a demotion to propose-only when the ratio exceeds 2x.

Key Safeguards

The author stresses that naive baselines are easily contaminated, prompting the addition of three protective mechanisms:

Learning freeze: Pauses the absorption of new events during a burn-rate anomaly, preventing the system from "learning failures as normal."
Asymmetric learning: Learns improvements quickly, but degrades slowly.
Absolute ceiling: Adds an absolute limit to catch slowly rising failure rates that might not trigger the ratio threshold.

Additional features include:

Per-agent and fleet-level kill switches.
A fleet() view for dashboards.
A dependency-free, MIT-licensed TypeScript implementation spanning a few hundred lines.
An interactive browser demo allowing users to inject drift and observe the breaker in action.
A longer article discussing the transfer of NOC operational experience to agent fleets (autonomy levels, error budgets, change freezes, etc.).

The author concludes by asking the community how they currently detect fleet-level degradation: per-run evaluations, baseline-relative signals, or more intelligent methods.

Original post →

Burn-Rate Circuit Breaker for Agent Fleets

More from coding & agent