Agent Control
Agent Control is an open-source control plane for AI agents. It enforces centralized policies at runtime, block, deny, steer, warn on each LLM output and tool call. We assess it for governance and guardrails across OpenAI, Move, Bedrock, and a growing set of agent frameworks.
Blurb
Control all AI agent behaviors at runtime with centralized policies. Write once and deploy everywhere.
Summary
What it is: A policy-driven step-level guardrail server for agent fleets. A @control decorator wraps LLM calls and tool execution, checking outputs against policies cached on the agent client, then responding with allow, deny, steer, or warn.
When to use: You need runtime governance, safety, and compliance across many agents. You want “write once, deploy everywhere” policy enforcement without redeploying agents.
When to skip: Single-agent apps with simple guardrails, where hard-coded checks suffice. You already run a gateway-side governance product that meets step-level needs.
Key features: Centralized Policy Enforcement, Control Store (mix guardrails from AWS Bedrock, Cisco AI Defense, NeMo, Galileo Luna), Audit Logs, Live policy updates without downtime, agent framework integrations (LangChain/LangGraph, Strands, Google ADK, OpenAI Agents SDK, AutoGen, CrewAI). Apache 2.0 licensed.
Details
Maintained by: Galileo (currently Galileo Technologies).
Guardrail ecosystem: Amazon Bedrock Guardrails, NVIDIA NeMo Guardrails, Galileo Luna-2, Azure AI Content Safety, Cisco AI Defense.