⚡ AgentBreaker

A real-time circuit breaker for AI agent loops.

AgentBreaker monitors token spend, iteration count, and cost velocity across multi-step LLM orchestration — and hard-stops runaway agents before they burn through budget or get stuck in infinite loops.

Most teams only set max_tokens on a single LLM call. AgentBreaker works at the orchestration layer — across multiple LLM calls, tool invocations, and sub-agent spawns. Observability tools tell you an agent went wrong after the fact. AgentBreaker stops it while it's happening.

---

The Problem

AI agents that loop, retry, or spawn sub-tasks can silently burn through API budgets in minutes. A max_tokens cap on one call doesn't catch an agent that makes 50 calls in a row. Teams find out when the bill arrives.

What AgentBreaker Does

• Tracks cost, token usage, iteration count, and elapsed time live, per agent run
• Evaluates every LLM call against a configurable rule engine
• Hard-stops the agent the instant a rule fires — not after the fact
• Surfaces exactly which rule fired and why, with full run history and cost trajectory

---

Features

• Composable rule engine — not hardcoded thresholds. Pick which rules apply per run:
  • Total Cost Limit
  • Max Iterations
  • Max Run Time
  • Spend Velocity (catches fast burns)
  • Stuck Loop Detector (same tool called repeatedly)
  • Cost Anomaly Spike (single call costs way more than average)
  • Long Run Warning (soft warning, doesn't stop execution)
• Live dashboard — real-time cost trajectory chart, run history, trip alerts with estimated savings
• REST API — start runs, poll status, fetch aggregate metrics
• Fully containerized — docker compose up and you have a working demo

---

Tech Stack

Layer	Tech
LLM Provider	Groq (Llama 3.3 70B)
Agent Framework	LangChain
Backend	FastAPI + SQLAlchemy + SQLite
Frontend	React + Recharts
Infra	Docker + Docker Compose
CI	GitHub Actions

---

Architecture

agentbreaker/

backend/ agent.py # Demo "runaway" research agent circuit_breaker.py # Core breaker — tracks state, evaluates rules rules.py # Rule engine + built-in rule library database.py # SQLAlchemy models, SQLite setup main.py # FastAPI app — endpoints + background run execution frontend/ src/App.js # Dashboard — run form, history, live detail view docker-compose.yml

---

Quick Start (Docker — recommended)

1. Clone the repo:

   git clone https://github.com/YOUR_USERNAME/agentbreaker.git
   cd agentbreaker

2. Get a free Groq API key at groq.com (no credit card required).

3. Create a .env file in the project root:

   echo "GROQ_API_KEY=your_key_here" > .env

4. Run it:

   docker compose up --build

5. Open http://localhost:3000

---

Local Development (without Docker)

Backend:

cd backend
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
uvicorn main:app --reload --port 8000

Frontend (separate terminal):

cd frontend
npm install
npm start

---

API Reference

Method	Endpoint	Description
POST	/runs	Start a new agent run
GET	/runs	List all runs
GET	/runs/{run_id}	Get a single run's live status
GET	/rules	List all available rules
GET	/metrics	Aggregate stats across all runs
DELETE	/runs	Clear all run history

---

Roadmap

• Phase 1 — Core breaker logic, demo agent, FastAPI backend, React dashboard
• Phase 2 — Composable rule engine, per-run rule configuration
• Phase 3 — Semantic loop detection, goal drift detection, anomaly baselines
• Phase 4 — LangChain callback handler for zero-friction integration, PyPI package

---

License

MIT