agentcontrol/agent-control

What it solves

Agent Control provides a centralized layer to enforce runtime guardrails for AI agents. It prevents risks such as prompt injections, PII (Personally Identifiable Information) leakage, and dangerous tool calls without requiring developers to hardcode safety logic into the agent's core code or prompts.

How it works

The system consists of a server and an SDK. Developers wrap model or tool calls with a @control() decorator. At runtime, the SDK sends inputs and outputs to the Agent Control server, which evaluates them against a set of configurable rules (controls). If a rule is violated, the system can block the action and throw a ControlViolationError.

Who it’s for

It is designed for developers building AI agents using frameworks like LangChain, CrewAI, Google ADK, and AWS Strands who need a way to manage safety policies across multiple agents from a single dashboard or API.

Highlights

• Centralized Management: Define and update safety controls via a UI or API without redeploying agent code.
• Pluggable Evaluators: Includes built-in evaluators for regex, lists, JSON, and SQL, with support for custom evaluators.
• Framework Agnostic: Compatible with various agent frameworks and supports both Python and TypeScript SDKs.
• Observability: Integrated support for OpenTelemetry and custom event sinks to track control executions.