Agent Control

Agent Control is an open source control plane for governing AI agents at run time. Instead of hard coding guardrails into each agent, teams define centralized policies once and enforce them at every step of an agent workflow, validating LLM outputs and tool calls and updating the rules without taking agents offline. It is built for engineering, platform and security teams running fleets of agents in production. Developers add the SDK to a Python agent with an import and a control decorator on the functions to protect, such as LLM calls or tool invocations. Each decorated step is checked against its assigned policies, either on the Agent Control server or locally within the agent, and the control returns allow, deny, steer or warn for the SDK to act on. Policies are cached on the client for speed and reliability. A control store lets teams mix guardrails from Amazon Bedrock Guardrails, NVIDIA NeMo Guardrails, Galileo Luna, Azure AI Content Safety and Cisco AI Defense or add their own, and audit logs capture full traces showing which controls fired, when and why. It works with custom agents and with LangChain and LangGraph, Strands Agents, Google ADK, the OpenAI Agents SDK, AutoGen and CrewAI. Typical uses include blocking invalid database access, validating agent permissions, limiting access to sensitive tools and preventing leaks of sensitive data. Agent Control is free under the Apache 2.0 license and is designed to be self hosted in your own environment. No paid plan is published on the site. The project is maintained by Galileo, the AI evaluation and observability company.