The Pulse
Microsoft introduces enterprise policy controls for AI agents
Microsoft's agent controls service lets enterprises define and evaluate policies for AI agents through Azure. Agent hosts receive decisions to allow, deny or transform actions and content at configured points during execution.

Enterprises can use a new Azure service to set and evaluate policies for AI agents across frameworks, platforms and clouds. Microsoft introduced agent controls on November 18 as a managed service that evaluates an agent's proposed actions and content while it is running.
The service is built on the open-source Agent Control Specification (ACS) and Agent Hooks. Microsoft says it is intended to give organizations centralized runtime governance without requiring application teams to operate their own policy-evaluation infrastructure.
Decisions during execution
Agent controls stores policies as Azure resources and evaluates them through a hosted API. At configured points in an agent's execution, the agent host submits an action or content for evaluation, along with runtime context.
The service returns an allow, deny or transform decision for the host to enforce. An allow decision directs the host to proceed; a deny decision blocks the action; and a transform decision calls for a change to be applied and validated before proceeding.
Each evaluation also produces a receipt identifying the policies and versions used. Applications record those receipts alongside execution outcomes, allowing teams to trace an action back to the policy decision.
Rules, classifiers and judges
The policies can combine Azure AI Content Safety classifiers, large language model judges and deterministic rules. Classifiers detect supported risk categories, while judges assess content against criteria written by the organization.
Rego rules can combine those results with the request context to produce a verdict using custom logic. The company's examples include transaction limits, restrictions on tool use and rules about approved destinations for data.
Other examples pair a classifier with a rule to check for personal information in external responses, or use a language model judge and a rule to assess whether a customer commitment needs review. The checks can address an agent's actions as well as the content it produces.
How the open components fit
Agent Hooks defines intervention points across an agent's lifecycle, including startup, input, model calls, tool calls, output and shutdown. At those points, the runtime pauses and requests a decision.
ACS specifies the request-and-decision contract. It gathers evidence through components such as classifiers, model judges, business rules or external APIs, then passes that evidence to a policy engine for a decision.
In a banking example, Microsoft describes a policy that sends transfers over $10,000 to human review. A customer in the example requests a $25,000 transfer, prompting the host to submit the proposed tool call for evaluation before it runs.
The same example uses other checks at different points. At input, a judge assesses a customer's message against a financial-hardship rubric; at output, a classifier checks a proposed response for personal information, with the policy configured to deny it or transform it to redact the information.
Source
Companies in this story
Related stories
- Microsoft and partners start enterprise AI transparency initiative
- FlowForma launches AI agent monitoring and control hub
- Microsoft makes Copilot Home generally available as Autopilot rolls out
- Microsoft Copilot SDK 2.0 supports shared agent sessions
- Malicious Microsoft Agent Framework releases removed from PyPI