Microsoft has introduced an open-source standard aimed at giving developers and companies more control over the behavior of AI agents.
The new framework, called Agent Control Specification (ACS), is designed to help organizations manage how AI agents operate across applications, workflows and environments.
As companies deploy increasingly capable AI systems, ensuring those agents remain within defined boundaries has become a growing concern.
Policy-based controls
ACS allows development, security and compliance teams to create policies that govern agent behavior.
Those policies can specify:
- Actions an agent is allowed to perform.
- Tasks it is prohibited from carrying out.
- Situations that require human approval.
- Information that should be logged for auditing.
The policies are checked at several stages while an agent performs a task.
Multiple checkpoints
Microsoft said guardrails can be applied at different points in an agent’s workflow, including:
- Before the agent receives input.
- Before it uses a tool.
- After a tool returns results.
- Before a response is sent to a user.
Depending on the policy, the system can:
- Allow an action.
- Block it.
- Remove sensitive information.
- Request approval from a human.
Replacing fragmented controls
Developers currently rely on a mix of techniques to manage AI behavior.
These include:
- System prompts.
- Application-level checks.
- Input and output classifiers.
While those approaches work, Microsoft said they can create fragmented systems that are difficult to audit and reuse across different frameworks.
ACS is intended to provide a common governance layer.
Additional safeguards
The specification also supports:
- Input and output classifiers.
- Large language models acting as policy “judges.”
- Checks on tool selection and tool calls.
- Input accuracy verification.
- Monitoring how outputs are used.
- Response validation.
Policies can be stored in a single file, allowing security rules to move with agents across different environments.
Broad framework support
Microsoft is releasing ACS as a software development kit with plug-ins for several AI frameworks and tools.
Supported platforms include:
- LangChain
- OpenAI Agents SDK
- Anthropic Agents SDK
- AutoGen
- CrewAI
- Semantic Kernel
- Microsoft.Extensions.AI
- MCP tools
TL;DR:
Microsoft has launched Agent Control Specification, an open-source standard designed to help developers manage AI agents. The framework lets organizations define policies governing what agents can do and when human approval is required.
AI summary:
- Microsoft introduced the open-source Agent Control Specification.
- ACS lets companies define rules for AI agent behavior.
- Policies can block actions, redact information or require human approval.
- The framework adds checkpoints throughout an agent’s workflow.
- ACS supports LangChain, OpenAI Agents SDK, Anthropic Agents SDK and other platforms.
