Microsoft’s Agent Controls Put Enterprise Agents on a Leash

Human Verified

Share to: , and

A control room with a wall dashboard of glowing connected agent nodes and a small robot

Microsoft has introduced Agent Controls, a managed Azure service that checks AI agent behaviour against company policy at runtime. The service evaluates each proposed action and returns allow, deny or transform, plus a receipt naming the policies applied. It is built for enterprises running agents across frameworks and clouds.

How It Works

Agent Controls rests on two open primitives Microsoft open-sourced: Agent Hooks and the Agent Control Specification. Hooks mark intervention points across an agent’s life, from startup and input through model calls, tool calls and output to shutdown. At each point the runtime pauses and asks for a decision.

The specification defines how each decision is made. Safety classifiers, LLM judges and deterministic Rego rules gather evidence, then a policy engine combines the results into a verdict. A policy written once should work everywhere, whatever framework or runtime runs the agent.

Three evaluation methods combine inside one policy. Azure AI Content Safety classifiers detect risk categories, LLM judges score content against written criteria, and Rego rules apply business logic such as spending limits. The service stores policies as Azure resources and evaluates them through a hosted API.

A Bank Agent, Checked

Microsoft’s worked example is a bank support agent that can look up accounts, move money and answer questions. A customer asks for a $25,000 transfer, and the agent prepares the call. At the tool-call hook, the host pauses and sends the proposed transaction for evaluation.

A deterministic rule caps transfers at $10,000 without human review, so the $25,000 request is denied. An LLM judge spots financial hardship in a customer’s message and routes them to a specialist instead of upselling. A PII classifier blocks personal data from leaving in an external response.

Each decision comes with a receipt naming the policies and versions applied. Teams record receipts beside execution outcomes, so every action traces back to its policy decision. That audit trail is the point: security teams can answer whether their agent fleet is actually governed.

Our Read

The honest limit is that Agent Controls decides, it does not enforce. The host applies each verdict, so the service is only as strong as the host’s wiring. Microsoft says so plainly: each evaluation returns a decision for the agent host to enforce.

Even so, the open standard is the real news. A policy written once, portable across frameworks and clouds, is what enterprises need as agents spread. Microsoft is already putting open agents inside Teams and Azure, and Agent Controls is the governance layer that stack was missing.

Why It Matters

Enterprises are moving from agent experiments to agent fleets, and governance has lagged behind. A shared policy engine means one rule can cover agents built by different teams on different frameworks. That is the difference between a policy document and a policy that actually runs.

Background

Agent Controls is the shared policy engine behind two Microsoft products. Foundry brings the controls into the developer workflow, attaching guardrails to agents as they are built. Agent 365 brings them into the governance workflow, where IT and security set policy once and see posture across the estate.

The service is in public preview, with no customers, pricing or general-availability date announced. Microsoft expanded its Azure AI pact with NVIDIA at Ignite, and Agent Controls is the control plane for that expansion. Teams can test it now with policy templates and the preview API.

Source: Microsoft’s announcement

Share to: , and

Faheem R. Khattak

Senior automation engineer and AI marketing journo covering LLMs, algorithms, analytics, core updates, Bing search, and technical SEO.

0 Comments

Leave a Reply

Optional Take Tag:

🛡️ Spam protection: Honeypot + Time-Trap Active

This site uses Akismet to reduce spam. Learn how your comment data is processed.