AgentBrake — The circuit breaker for LLM agents in production

The problem

Your agents are running unsupervised.

Observability tells you it happened. AgentBrake stops it from happening.

Infinite Loops

Agent retries the same failing call 50 times before giving up.

Budget Blowups

A runaway agent burns $200 in tokens overnight.

Privilege Escalation

Agent calls delete_database because a user prompt-injected it.

The solution

3 lines of code. Full protection.

Wrap your tool-dispatch function. If the agent loops, blows the budget, or calls something off the allowlist, it stops — before the tool runs.

agent.py

import agentbrake

agentbrake.init(allowed_tools=["search", "read_file"], budget_usd=5.0)

@agentbrake.guard()
def call_tool(name: str, args: dict):
    return my_tools[name](**args)

That's it. On a trip, call_tool raises AgentBrakeInterrupt 🛑 instead of executing — catch it and handle the LOOP, BUDGET, or ESCALATION reason however you like.

How it works

Detect → Interrupt → Validate

Every tool call passes through the SDK before it ever runs.

01

Detect

SDK monitors every tool call for loops, budget overruns, and unauthorized tools.

02

Interrupt

Stops the agent mid-run before damage happens. Not after.

03

Validate

Sends a link to Slack/browser. Human approves or kills in one click.

Comparison

Enforcement, not just observability.

The others show you what your agent did. AgentBrake is the brake pedal — it acts before the tool executes.

Tool	Approach	When it acts	Self-hosted
AgentBrake Best	Enforcement (circuit breaker)	Before damage (mid-run)	✓ Yes (MIT)
LangSmith	Observability + guardrails	After + during	✗ No (cloud)
Helicone	Observability + caching	After	✓ Yes (open core)
AgentOps	Observability + replay	After	✗ No (cloud)

We don't compete with these — we complement them. Run AgentBrake as your last line of defense before the tool actually executes.