AI Agents Get Guardrails With New Microsoft Spec

Enterprises deploying AI agents across various applications now face the challenge of ensuring these agents operate as intended. Microsoft is addressing this with its open-source Agent Control Specification (ACS), providing developers a more consistent and granular method for controlling AI agent actions.

ACS enables development, compliance, and security teams to establish specific policies for AI agents. These rules dictate permissible actions, prohibited behaviors, required human approvals, and necessary evidence logging for reviews. The specification integrates these controls into a common governance layer, checking policies at multiple interception points within an agent's workflow.

This standard aims to overcome the limitations of current ad-hoc methods, such as system prompts or custom code checks, which often result in fragmented and hard-to-audit controls. ACS can be applied before input, before tool calls, after tool results, and before final user responses, allowing actions to be allowed, blocked, redacted, or approved by humans.

The specification supports inserting classifiers for input/output categorization, using LLMs as policy judges, and defining logic for tool calls, input accuracy, and response validation. Because policies are single files, they can be bundled with agents, ensuring security protocols travel with the agent across different frameworks and environments. ACS is shipping as an SDK with plugins for numerous popular AI frameworks like LangChain and AutoGen.