Microsoft unveils a simpler way to keep AI agents on task

As companies rush to deploy AI agents across their operations, a growing headache has emerged: how to make sure these autonomous systems actually follow the rules when they’re working in different apps and environments. Microsoft thinks it has an answer with a new open-source framework called the Agent Control Specification, or ACS.

The idea is straightforward. Instead of cobbling together instructions in system prompts, custom code checks, or classifiers scattered across a project, ACS lets developers, compliance officers, and security teams write clear policy files. These files spell out what an agent can do, what it must avoid, when a human needs to sign off, and what data gets logged for later audits. The system checks those rules at multiple points during the agent’s workflow—before it receives input, before it calls a tool, after the tool responds, and before the final reply goes out. Depending on the policy, the agent might proceed, get blocked, redact sensitive details, or pause for approval.

The specification is designed to fix the fragmentation that plagues many current approaches. Teams often end up with a patchwork of controls that are tough to audit and harder to reuse across different tools or platforms. ACS packages those rules into single files that travel with the agent, so security policies stay consistent whether the agent is running on LangChain, the OpenAI Agents SDK, Anthropic’s SDK, AutoGen, CrewAI, Semantic Kernel, or other frameworks.

Microsoft is shipping ACS as an SDK with plugins for those systems, plus support for MCP tools and more. The goal is to give developers a common governance layer that works across the board, without forcing them to reinvent the wheel for every new agent or environment.