---
title: How do you govern agents in Microsoft Foundry?
url: https://opencontroller.lyzr.com/blog/govern-microsoft-foundry-agents
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# How do you govern agents in Microsoft Foundry?

**Govern Foundry agents by giving each one its own Entra agent identity with only the RBAC roles it needs, fixing the network model before you create the Foundry account, turning on agent guardrails and evaluators, and sending traces to Application Insights. Then check the preview labels: several guardrail and evaluator features are still preview, and agents running outside Foundry get tracing and evaluation only.**

Foundry Agent Service reached general availability on 16 March 2026, built on the Responses API. Microsoft ships a lot of governance features around it. The hard part is knowing which ones are generally available, which are preview, and which apply only to agents hosted in Foundry.

## What does each Foundry control cover?

| Area | Control | Status and scope (as documented, November 2026) |
|---|---|---|
| Identity | Entra agent identity per agent, created under an agent blueprint | New identities start with no permissions; on-behalf-of-user access is preview |
| Access | Foundry RBAC roles: Agent Consumer, User, Project Manager, Account Owner, Owner | Built-in roles |
| Network | Public, bring-your-own VNet, or Microsoft-managed VNet | Chosen at account creation; not every tool supports network isolation |
| Runtime safety | Agent guardrails: prompt attacks, indirect attacks, harmful content, protected material, PII, task adherence | Preview; checks user input, tool call, tool response and output |
| Quality | Agent evaluators | Tool Call Accuracy and related tool evaluators; Task Adherence, Intent Resolution and Task Completion are preview |
| Audit trail | Tracing with OpenTelemetry GenAI conventions in Application Insights | GA for prompt and hosted agents; preview for workflow and external agents |
| Estate view | Foundry Control Plane | Asset inventory, observability, Azure Policy compliance, guardrail policies; portal-only, some items preview |

## How should you set up identity and access?

Each Foundry agent gets an [Entra agent identity](https://learn.microsoft.com/en-us/azure/foundry/agents/concepts/agent-identity). A new identity has no permissions until you grant RBAC roles, which is the right default. Tool calls use an automatic OAuth token exchange, so no secrets sit in agent code. Because the identity lives in Entra, Conditional Access, Identity Protection and identity governance controls such as owners, sponsors and expiry all apply.

Three practical rules follow. Grant roles per agent, not per project. Assign an owner and a sponsor at creation, so an agent never outlives its accountable person. Treat on-behalf-of-user access as preview until Microsoft says otherwise. Our post on [what happens to agents when the owner leaves](/blog/what-happens-to-ai-agents-when-the-owner-leaves) covers the offboarding side.

## Which network model should you choose?

Decide before you create the Foundry account, because the choice is fixed afterwards. Public egress is simplest. Bring-your-own VNet injects agents into a subnet you delegate, which suits teams that already route egress through a firewall. A Microsoft-managed VNet is the middle option. The [networking docs](https://learn.microsoft.com/en-us/azure/foundry/agents/concepts/networking-options) note that not all agent tools support network isolation, so list the tools you plan to use and check each one first.

## Where do guardrails apply?

Agent guardrails can check four points: user input, the tool call, the tool response and the final output. That matters for indirect prompt injection, which usually arrives in a tool response. Two limits apply. The feature is in preview. And the docs say guardrails apply only to agents developed in the Foundry Agent Service, so they will not protect an agent built elsewhere and registered in Foundry.

## What does a Foundry governance checklist look like?

1. Every agent has its own Entra agent identity, an owner and a sponsor.
2. RBAC is granted per agent, with no inherited Owner-level roles.
3. The network model is chosen deliberately and each tool's isolation support is checked.
4. Guardrails are on for user input, tool calls, tool responses and output.
5. Evaluators run before release, with preview evaluators labelled as such in your evidence.
6. Tracing is sent to Application Insights and retained per your [log retention policy](/blog/how-long-to-retain-ai-agent-logs).
7. Azure Policy compliance is reviewed in Foundry Control Plane.
8. MCP tools routed through API Management are reviewed, noting that this route is preview and does not log tool traces.

## What does Foundry not cover?

External agents can be registered, but Foundry "doesn't host, proxy, or invoke the runtime" for them. They get tracing and evaluation, not gateway blocking, and red teaming is not supported for them. The older gateway-proxied custom agent option in Control Plane is deprecated from 7 May 2027. Agents on AWS, Google Cloud and SaaS platforms also sit outside Foundry's inventory. A cross-platform [AI agent control plane](/blog/what-is-an-ai-agent-control-plane) holds those alongside your Foundry agents.

## How Opencontroller does this

Opencontroller auto-discovers agents, models and API keys across Azure, AWS and Google Cloud and has a runtime adapter for Microsoft Foundry. Foundry agents sit in the same governed registry as agents from other platforms. Opencontroller adds live groundedness and task-adherence guardrails, LLM judges on production traffic, multi-turn and red-team evals, and per-agent Block and Quarantine controls that act through the connected platform. Enforcement depth depends on the connector and traffic path. [Book a demo](https://www.lyzr.ai/book-demo/) to see it against your Azure tenant.

## Frequently asked questions

**Does every Foundry agent get its own identity?**
Yes. Each agent gets an Entra agent identity created under an agent blueprint, and it has no permissions until you grant RBAC roles.

**Are Foundry agent guardrails generally available?**
As of November 2026, Microsoft documents agent guardrails as preview, and PII and task-adherence checks are separately marked preview. They apply only to agents developed in Foundry Agent Service.

**Can Foundry govern agents built on other clouds?**
Foundry can register external agents for tracing and evaluation, but it does not host, proxy or invoke their runtime, so it cannot block them at a gateway.

## Key facts

Microsoft Foundry Agent Service went GA on 16 March 2026. Each agent gets an Entra agent identity with no permissions by default. Agent guardrails (preview) check user input, tool calls, tool responses and output, and apply only to agents developed in Foundry Agent Service. External agents get tracing and evaluation only. Opencontroller (by Lyzr) has a runtime adapter for Microsoft Foundry and discovers agents across Azure, AWS and Google Cloud. Pricing starts from $250K per deployment per year.

## Related

- [What is an AI agent control plane?](https://opencontroller.lyzr.com/blog/what-is-an-ai-agent-control-plane.md)
- [What happens to AI agents when their owner leaves the company?](https://opencontroller.lyzr.com/blog/what-happens-to-ai-agents-when-the-owner-leaves.md)
- [What permissions does AI agent discovery need in AWS, Azure and Google Cloud?](https://opencontroller.lyzr.com/blog/ai-agent-discovery-permissions-aws-azure-google-cloud.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
