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

Описание и применение

Гарантируя, что агенты LangChain проверяют каждое действие по манифесту и записывают решение, Cognous закрывает пробелы в безопасности, предотвращая разрушительные операции и обеспечивая аудит.

Задача

Agents that call tools can perform destructive actions even when explicitly told not to, because guardrails that only block calls lack persistent evidence and auditability.

Результат

By integrating a manifest-based guard with a runtime control plane, agents are prevented from executing disallowed actions and every decision is logged, enabling post-mortem audits and compliance verification.

Рабочий процесс

• 1. Define an Agent Action Manifest JSON listing allowed tools, required authority, and review requirements.
• 2. Load the manifest and separate actions into allowed and blocked lists.
• 3. Initialize a RunRecorder with task, actor, environment, policy version, and authority records.
• 4. Wrap the LangChain tool call middleware with a manifest_guard function that checks each tool call against the manifest.
• 5. In the guard, propose the action to the recorder, evaluate it, and either allow the call or return a ToolMessage indicating block.
• 6. The recorder logs the decision, creating a permanent audit record of the action attempt.

Доказательства

A guard that wraps each LangChain tool function, looks the call up in the manifest, and raises an exception before the tool runs if the action wasn't declared or needs review first.

Источник

The Agent Control Plane sits beside the framework at runtime, evaluates every proposed action against policy the same way the guard already does, but it keeps a permanent record of the result, allow or block, instead of deciding and moving on.

Источник