Prove the Guardrails Worked: Turning a LangChain Agent's Run Into a Replay Bundle

A CISO inquired about the behavior of an AI agent in a specific run, asking for details about the actual actions taken during that particular execution. Directly searching through every server's application logs to answer such a question is not an efficient solution. Cognous' Open Control Stack was designed to address this issue, providing a way to answer the question in minutes rather than days.

The Open Control Stack organizes AI agent governance into four layers: declaring what an agent may do, controlling what it actually does at runtime, replaying any specific run after the fact, and packaging the result as evidence for review. Wiring the Agent Action Manifest into LangChain covered the first layer, while wiring LangChain into the Agent Control Plane covered the second layer.

The Control Plane records each decision made by the agent, answering "what did the policy decide." However, to provide a record that can be handed to an auditor, the RunRecorder export from the Control Plane's object model needs to be turned into something portable, independently verifiable, and safe to share outside the team that built the agent. This is where the Replay layer comes in.

The Replay layer defines its own schema for a portable replay bundle called AgentReplayBundle, which exists independently from the Control Plane. However, there is a mismatch between the Control Plane's export and the arb's schema, requiring a mapping step to convert the Control Plane's internal snapshot into the public interchange format necessary for validation, redaction, signing, and verification.