AGENTICONTROL V0 MVP

Control plane for AI agents with synchronous blocking and asynchronous logging.

Overview

AGENTICONTROL provides:

• Synchronous Blocking: Zero-latency policy checks and loop detection
• Asynchronous Logging: Non-blocking trace event ingestion
• Terminal Viewer: Real-time structured trace output for local debugging
• Cloud Backend: Scalable trace storage and analytics

Key Features

• 🛑 Policy V0 Checks: Block dangerous operations (SQL DROP, DELETE, PII patterns)
• 🔄 Loop Detection: Detect and halt infinite loops (Rule A & B)
• 💰 Cost Monitoring: Track token usage and estimated costs
• 📊 Terminal Viewer: Beautiful real-time trace visualization with rich
• ☁️ Cloud Uplink: Async trace ingestion to Supabase

Installation

pip install agenticontrol

Quick Start

Installation

pip install agenticontrol

Basic Usage with LangChain

from langchain.agents import initialize_agent, AgentType
from langchain.llms import OpenAI
from agenticontrol.hooks.langchain_handler import AgenticontrolCallbackHandler

# Initialize the callback handler
handler = AgenticontrolCallbackHandler(
    api_url="http://localhost:8000/api/v1/ingest/trace",  # Your backend URL
    api_key=None,  # Optional API key
    enable_blocking=True,  # Enable Policy V0 and Loop Detection
    enable_logging=True,  # Enable async trace logging
)

# Use with LangChain agent
agent = initialize_agent(
    tools=tools,
    llm=llm,
    agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
    callbacks=[handler],
    verbose=True
)

# Run agent - blocking checks happen automatically
try:
    result = agent.run("Your query here")
    print(f"Result: {result}")
except PolicyViolationError as e:
    print(f"Agent blocked: {e}")

# Cleanup
import asyncio
asyncio.run(handler.flush())
asyncio.run(handler.close())

Local-Only Mode (No Cloud Backend)

from agenticontrol.hooks.langchain_handler import AgenticontrolCallbackHandler

# Local-only mode: blocking checks work, but no cloud logging
handler = AgenticontrolCallbackHandler(
    api_url=None,  # Local-only mode
    enable_blocking=True,
    enable_logging=False
)

Terminal Viewer

The terminal viewer provides real-time trace visualization:

from agenticontrol.local.viewer import TraceViewer, get_viewer

# Get the viewer instance
viewer = get_viewer()

# Events are automatically displayed when using the callback handler
# The viewer shows:
# - 💭 LLM prompts and responses
# - 🛠 Tool calls and outputs
# - 🛑 Blocked operations (Policy V0, Loop Detection)
# - 💰 Cost tracking

Examples

See the examples/ directory for complete examples:

• basic_usage.py - Basic LangChain integration
• loop_detection_demo.py - Loop detection demonstration
• policy_v0_demo.py - Policy V0 checks demonstration

Features

🛑 Policy V0 Checks

Blocks dangerous operations synchronously (zero latency):

• SQL commands: DROP TABLE, DELETE FROM
• File system: rm -rf, dangerous paths
• PII patterns: Email addresses, SSNs, etc.

# Automatically blocks dangerous operations
agent.run("DROP TABLE users")  # Raises PolicyV0ViolationError

🔄 Loop Detection

Detects and halts infinite loops (Rule A & B):

• Rule A: Identical LLM prompts repeated
• Rule B: Identical tool calls repeated

# Blocks after 5 identical tool calls
for i in range(10):
    agent.run("search for 'test'")  # Blocks on 6th identical call

💰 Cost Monitoring

Tracks token usage and estimated costs:

• Accumulates tokens per run
• Calculates USD cost based on model rates
• Updates in real-time

📊 Terminal Viewer

Beautiful real-time trace visualization:

• Tree-like structure showing agent execution
• Color-coded events (LLM, tools, errors)
• Prominent display of blocked operations

☁️ Cloud Uplink

Async trace ingestion to Supabase:

• Non-blocking HTTP calls
• Batched uploads for efficiency
• Never delays agent execution

Architecture

Synchronous Blocking (Zero Latency)

• Policy V0 checks run in-memory
• Loop detection runs synchronously
• Raises PolicyViolationError to halt execution immediately

Asynchronous Logging (Non-Blocking)

• Trace events sent via async HTTP
• Batched uploads for efficiency
• Never blocks agent execution

API Reference

AgenticontrolCallbackHandler

Main callback handler for LangChain integration.

Parameters:

• api_url (str, optional): Backend API URL for trace ingestion
• api_key (str, optional): API key for authentication
• run_id (UUID, optional): Unique identifier for this run (auto-generated)
• risk_engine (RiskEngine, optional): Custom risk engine instance
• enable_blocking (bool): Enable synchronous blocking checks (default: True)
• enable_logging (bool): Enable async trace logging (default: True)

Methods:

• flush(): Flush pending events to cloud (async)
• close(): Close client connections (async)

Exceptions

• PolicyViolationError: Base exception for all policy violations
• PolicyV0ViolationError: Raised when Policy V0 check fails
• LoopDetectedError: Raised when loop detection triggers
• CostThresholdExceededError: Raised when cost threshold exceeded

Development

Setup

# Clone repository
git clone https://github.com/yourorg/agenticontrol.git
cd agenticontrol

# Install in development mode
pip install -e ".[dev]"

# Run tests
pytest

# Run with coverage
pytest --cov=src/agenticontrol --cov-report=html

# Format code
black src tests examples

# Lint code
ruff check src tests examples

# Type checking
mypy src

Running Examples

# Basic usage
python examples/basic_usage.py

# Loop detection demo
python examples/loop_detection_demo.py

# Policy V0 demo
python examples/policy_v0_demo.py

Running Backend

cd backend
python run_server.py
# Or: python -m uvicorn backend.app.main:app --reload

Project Structure

src/agenticontrol/
├── models.py              # Pydantic models (TraceEvent, RunMetadata, RiskResult)
├── risk_engine.py         # Synchronous blocking checks
├── client.py              # Async HTTP client
├── exceptions.py          # PolicyViolationError hierarchy
└── hooks/
    └── langchain_handler.py  # LangChain integration

local/
└── viewer.py              # Terminal trace viewer

backend/
└── app/                   # FastAPI backend (separate repo)

Schema Versioning

The TraceEvent model includes a schema_version field for backward compatibility. Current version: 1.0.0

License

MIT

Contributing

See CONTRIBUTING.md (coming soon)