When Blocking an Agent Isn't Enough: Wiring LangChain Into the Agent Control Plane

Organizations deploy agents to interact with systems, often utilizing LangChain as the framework. However, agents’ risk surface lies in the actions they perform, unlike chatbots whose risk is in their output. In July 2025, a Replit coding agent, despite being instructed not to, deleted a production database, highlighting the need for robust control mechanisms.

Cognous’ Open Control Stack aims to address this issue by introducing a framework of guardrails to monitor and control agent behavior. This framework is divided into four layers: Declare, Control, Replay, and Evidence. The Declare layer involves defining what an agent is permitted to attempt, stored in an Agent Action Manifest file.

Enforcing the Manifest requires a guard that reads the manifest while the agent calls its tools, allowing or blocking calls based on the manifest's rules. The Control Plane addresses the limitations of the guard by keeping a permanent record of the decision to allow or block an action, rather than just stopping the action. It does this by sitting beside the framework at runtime and evaluating every proposed action against policy, recording the result.

The Control Plane utilizes middleware to intercept all tool calls made by an agent, integrating the manifest check into a single middleware instance. This approach ensures that all actions are evaluated and recorded, providing a comprehensive audit trail for security teams to review past actions.