agent-fuse 0.2.0

A Circuit Breaker for AI Agents - Fail Safe by Default

AgentFuse

The Circuit Breaker for AI Agents.

Stop infinite loops and wallet-draining bugs before they hit the API.

pip install agent-fuse

---

The Problem

You write an Agent. You give it a loop. You go to sleep.

You wake up to a $500 OpenAI bill because your agent got stuck in a while True loop trying to fix a typo.

This happens more than you think:

• AutoGen agents retrying failed tool calls
• LangChain chains with bad exit conditions
• RAG pipelines re-embedding the same documents
• Any agent with access to gpt-4 and poor error handling

The Solution

Agent Fuse is a local circuit breaker for your LLM calls. It sits between your code and OpenAI, enforcing:

1. Hard Budget Limits - "Max $2.00 per session"
2. Pre-flight Checks - Block calls before they're made
3. Loop Detection - Stop agents stuck in retry loops
4. Fail-Safe Architecture - Won't break your app if the guard fails

Zero latency. Zero external dependencies. Just SQLite.

---

Quick Start (30 seconds)

1. Install

pip install agent-fuse

Or with optional integrations:

pip install agent-fuse[openai]      # OpenAI support
pip install agent-fuse[langchain]   # LangChain support
pip install agent-fuse[all]         # Everything

2. Protect Your Agent

Just change your import. Agent Fuse is a drop-in replacement.

# Before
from openai import OpenAI
client = OpenAI()

# After
from agent_fuse import init
from agent_fuse.integrations import AgentFuseOpenAI as OpenAI

init(budget=5.00)  # Max $5.00 for this session
client = OpenAI()  # Works exactly the same!

That's it. If your agent tries to exceed $5.00, Agent Fuse raises SentinelBudgetExceeded and your wallet stays intact.

---

Features

Drop-in OpenAI Replacement

from agent_fuse import init
from agent_fuse.integrations import AgentFuseOpenAI

init(budget=10.00)
client = AgentFuseOpenAI()

# Use exactly like openai.OpenAI
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello!"}],
    stream=True  # Streaming supported!
)

LangChain Integration

from agent_fuse import init
from agent_fuse.integrations import AgentFuseCallbackHandler
from langchain_openai import ChatOpenAI

init(budget=5.00)
handler = AgentFuseCallbackHandler()

llm = ChatOpenAI(model="gpt-4o", callbacks=[handler])
response = llm.invoke("What is 2+2?")  # Protected!

Manual Pre/Post Flight

For custom integrations or non-OpenAI models:

from agent_fuse import init, pre_flight, post_flight, monitor

init(budget=5.00)

# Before your LLM call
pre_flight("gpt-4o", estimated_input_tokens=1000)

# ... make your API call ...

# After your LLM call (with actual usage)
post_flight("gpt-4o", input_tokens=950, output_tokens=500)

# Check your spend
stats = monitor()
print(f"Spent: ${stats.total_spend_usd:.4f}")
print(f"Remaining: ${stats.budget_remaining_usd:.2f}")

Loop Detection

Prevent agents from getting stuck in retry loops. AgentFuse tracks tool actions (not LLM outputs) and raises an error if the same action is repeated too many times.

from agent_fuse import init, check_loop, SentinelLoopError

init(budget=5.00, loop_threshold=5)  # Error after 5 identical calls

# In your agent's tool execution:
def execute_tool(tool_name: str, args: dict):
    try:
        check_loop(tool_name, args)  # Raises after 5 identical calls
    except SentinelLoopError as e:
        print(f"Agent stuck! {e}")
        print(f"Signature: {e.signature}")  # For debugging
        raise

    return tools[tool_name](**args)

Or use the decorator:

from agent_fuse import loop_guard

@loop_guard()
def search_web(query: str) -> list[str]:
    return api.search(query)

# Works up to 5 times with same args
search_web("python tutorial")  # Call 1
search_web("python tutorial")  # Call 2
# ...
search_web("python tutorial")  # Call 6 - Raises SentinelLoopError!

# Different args = separate counter
search_web("rust tutorial")    # Call 1 (new counter)

Why track actions, not thoughts?

Agents think differently each iteration ("Let me try again", "One more time...") but when they're stuck, they do the same thing. By hashing tool calls (name + args), we catch loops without false positives from varying LLM outputs.

Fail-Safe Mode

By default, Agent Fuse prioritizes safety - if the database fails, your agent stops.

For high-availability scenarios, you can flip this:

from agent_fuse import init

# Safety mode (default): Block agent if Agent Fuse fails
init(budget=10.00, fail_safe=True)

# Availability mode: Log warning but let agent continue
init(budget=10.00, fail_safe=False)

---

Configuration

All settings can be configured via environment variables:

Variable	Default	Description
AGENTFUSE_BUDGET	1.00	Maximum spend in USD
AGENTFUSE_FAIL_SAFE	True	Block on errors (safety) or continue (availability)
AGENTFUSE_DB_PATH	~/.agent_fuse/guard_v1.db	SQLite database location
AGENTFUSE_MAX_RETRIES	3	Retry attempts for DB operations
AGENTFUSE_LOOP_THRESHOLD	5	Identical tool calls before loop error
AGENTFUSE_LOOP_DETECTION_ENABLED	True	Enable/disable loop detection

Or configure programmatically:

from agent_fuse import init

init(
    budget=10.00,
    fail_safe=True,
    session_id="my-agent-run-123",
    db_path="/tmp/agent_fuse.db",
    loop_threshold=5,
    loop_detection_enabled=True
)

---

Supported Models

Agent Fuse includes pricing for 25+ models:

OpenAI: gpt-4o, gpt-4o-mini, gpt-4-turbo, gpt-4, gpt-3.5-turbo, o1, o1-mini

Anthropic: claude-3-opus, claude-3-sonnet, claude-3-haiku, claude-3.5-sonnet, claude-3.5-haiku

Unknown models fall back to conservative pricing estimates.

---

Architecture

Your Agent
    |
    v
+------------------+
| AgentFuseOpenAI  |  <- Drop-in replacement
+--------+---------+
         |
    +----+----+
    v         v
Pre-flight  Post-flight
 (check)     (record)
    |         |
    +----+----+
         v
+------------------+
| SQLite (WAL)     |  <- Thread-safe, concurrent writes
+------------------+
         |
         v
   ~/.agent_fuse/
   guard_v1.db

Why SQLite?

• Zero configuration
• WAL mode handles concurrent agents
• No network dependency
• Works offline

---

API Reference

init(budget, fail_safe, session_id, db_path, loop_threshold, loop_detection_enabled)

Initialize Agent Fuse. Call once at startup.

pre_flight(model, estimated_input_tokens, ...)

Check if a call would exceed budget. Raises SentinelBudgetExceeded if over.

post_flight(model, input_tokens, output_tokens)

Record actual usage after a call completes.

monitor()

Returns UsageStats with current spend, remaining budget, and call counts.

guard(model)

Decorator for wrapping functions with pre-flight checks.

check_loop(tool_name, args, session_id)

Check if a tool call would create a loop. Raises SentinelLoopError if threshold exceeded.

loop_guard(tool_name)

Decorator for wrapping functions with loop detection.

---

Exceptions

from agent_fuse import (
    SentinelBudgetExceeded,  # Budget limit hit
    SentinelSystemError,     # DB failed (in fail_safe mode)
    SentinelLoopError,       # Loop detected (tool called too many times)
)

# SentinelLoopError attributes:
# - call_count: Number of times the call was made
# - signature: The tool call signature (tool_name|args)
# - pattern: Alias for signature

---

Development

# Clone
git clone https://github.com/agentfuse/agent-fuse
cd agent-fuse

# Install dev dependencies
python -m venv env
source env/bin/activate
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Run verification
python verify_phase1.py

---

Roadmap

• Budget limits
• OpenAI shim (sync + streaming)
• LangChain callback handler
• Fail-safe architecture
• Loop detection (repeated similar calls)
• Rate limiting (calls per minute)
• Multi-agent session tracking
• Dashboard / CLI stats viewer

---

Limitations

Transparency builds trust. Here's what AgentFuse does NOT do (yet):

• No rate limiting - It's a budget guard, not a rate limiter. Use OpenAI's built-in rate limits for that.
• Single-machine only - The SQLite DB is local. For distributed agents, you'd need a shared backend (PRs welcome!).
• Estimates can drift - Pre-flight estimates use heuristics. Actual costs are recorded post-flight, but a large call could slip through if estimates are low.

---

License

MIT

---

Contributing

PRs welcome. Please include tests.

---

Built with paranoia by developers who've seen $500 bills.