Microsoft Empowers Developers with Greater Control Over AI Agent Behavior

As AI agents advance in complexity, businesses aiming to deploy them in a variety of applications, workflows, and products face a pivotal challenge: guaranteeing that agents function as expected in different environments.

To tackle this challenge, Microsoft has unveiled an open-source standard called the Agent Control Specification (ACS), intended to provide developers with a more consistent and comprehensive approach for overseeing AI agent behavior.

This specification empowers developer, compliance, and security teams to create their own policies for agents. These regulations can specify which actions the agent is allowed to perform, which it should refrain from, situations that necessitate human approval, and what evidence should be documented for later evaluation. Policy files are assessed at multiple “interception points” as the agent executes tasks to ensure adherence to these guidelines.

The launch of this specification is timely, as developers are currently devising ad-hoc strategies to manage AI behaviors and actions, especially amid conversations concerning AI workflow failures caused by tool misuse or unintended actions resulting in major problems.

Currently, developers may articulate instructions in system prompts, integrate custom checks within the application code, or utilize classifiers to identify problematic inputs and outputs. While these approaches can be effective, they often lead to fragmented controls that are challenging to audit and difficult to replicate across various frameworks, interfaces, and systems.

ACS aims to unify these controls into a comprehensive governance framework. According to Microsoft, this specification can verify whether an agent complies with the established guidelines at various stages of its workflow—before receiving input, prior to invoking a tool, after a tool provides results, and before delivering the final response to the user. A policy can allow an action, forbid it, redact sensitive information, or require human oversight.

Developers also have the option to implement classifiers for inputs and outputs to categorize data, predict outcomes, or define how an agent should respond; integrate large language models (LLMs) with prompts to act as a “judge” for compliance; and apply logic for assessing tool calls, selecting tools, verifying input accuracy, managing output usage, and crafting responses.

Additionally, since these policies can be contained within single files, they can be bundled with agents, ensuring a security policy accompanies each agent across various frameworks and environments.

ACS is being released as an SDK featuring plugins for LangChain, the OpenAI Agents SDK, the Anthropic Agents SDK, AutoGen, CrewAI, Semantic Kernel, Microsoft.Extensions.AI, MCP tools, and more.