Example code demonstrates loading a manifest, distinguishing allowed and blocked actions, starting a RunRecorder, and implementing a manifest_guard that checks calls against the manifest and records outcomes.

Cognous proposes the Open Control Stack—Declare, Control, Replay, Evidence—as a four-layer structure to prevent unwanted agent actions and provide traceability.

Six months after execution, the system preserves a structured record of decisions and blocks, enhancing auditability and accountability beyond volatile logs.

The article shows scaling guardrails from decorators to LangChain middleware to intercept all tool calls in one place, rather than wrapping each function.

It is implemented by a manifest-aware guard and a runtime Control Plane that enforces policy and permanently records each decision as either allow or block for accountability.

The article warns that agents with tool access can act autonomously and cause harm, citing a July 2025 incident where a coding agent deleted a production database despite clear prohibitions.

Cognous promotes open-source resources—the Agent Action Manifest, Agent Control Plane, and Open Control Stack—as ready-made guardrails and audit trails for LangChain agents.

Runtime records explicitly show outcomes for three tool calls: an allowed action (added column), a blocked action (table_drop), and an escalated action (schema_rename_table) that requires human review.

Declaration is implemented via an Agent Action Manifest that lists tools, required authority, and which actions can run unattended versus those needing human approval.

LangChain middleware wrap_tool_call routes tool calls through a manifest guard, integrating manifest checks with runtime execution.

The Control Plane bridges naming gaps between manifest and policy, ensuring policy is consistently enforced across runs with a post-run export that captures decisions and blocks.