A woman monitors a secure cloud-based dashboard with AI tools, server connections, and access controls.
Microsoft used Ignite '26 to add enterprise controls to the Browser Automation tool in Foundry Agent Service. The tool itself isn't new. Microsoft's October 19 announcement covered GA support for prompt agents and hosted agents. The November 18 Foundry Blog post builds on that. It adds networking, credential and identity features, and most of them are previews. Anyone planning a production rollout should read the status labels carefully.

A woman monitors a secure cloud-based dashboard with AI tools, server connections, and access controls. GA versus preview: the scorecard​

Microsoft's post says the Browser Automation tool in Foundry Agent Service, powered by Playwright Workspaces as its headless browser infrastructure layer, is generally available. The listed capabilities do not all share that status.

CapabilityStatus in the announcement
Expanded MCP tool primitives with action batchingGA
Live View and Take controlGA
Traces and session recordingGA
Private websites via VNet injectionPreview
Bring-your-own browser proxyPreview
Secret managementPreview
Web Bot Auth (listed under CAPTCHA management)Preview
Profile reusePreview, "coming soon"

Microsoft also says preview capabilities are subject to change. It tells customers to check the documentation before using them in production.

What the GA core does​

A richer set of browser tools​

The post describes a larger set of MCP tools built for browser work. Step-by-step tools cover:

  • opening and reading pages
  • accessibility snapshots
  • clicking and typing
  • completing forms and selecting options
  • handling dialogs
  • managing tabs
  • uploading files
  • capturing screenshots, console messages and network requests

For multistep sequences, a run-code tool runs one Playwright script for the whole sequence. The alternative is a long chain of snapshot, click and type calls.

Microsoft's advice is:

  • Use batched scripts for repetitive flows.
  • Use step-by-step tools for unfamiliar pages and dialog-heavy flows, where checking the page at each step is worthwhile.

Scripts can run arbitrary code, so run-code is opt-in. Microsoft says to enable it only for reviewed scenarios. In practice, treat it like any other code-execution path. Review it, scope it, and don't turn it on by default.

Live View and Take control​

Live View shows the agent's browser in headed mode, in real time. Take control lets a person step in to complete a sign-in, give an approval or recover from an unexpected dialog. They then hand control back and the agent continues.

Microsoft's October GA guidance adds governance advice:

  • Require approval before sensitive or state-changing operations.
  • Limit agents to the browser tools a task needs.
  • Use least-privilege access with managed identities and Azure RBAC.
  • Treat web content as untrusted input.

Human oversight helps, but it doesn't make an automated workflow correct by itself. An operator who isn't watching Live View has nothing to approve.

Recording and audit trail​

Session recording is opt-in. Recordings go to your own storage account, so storage RBAC decides who can replay them. Microsoft says activity logs and billing observability are also available. The sample connection JSON in the post shows an enableRecording flag set to "false", so recording isn't on by default in that example.

The same sample has placeholder values and an apparent formatting defect, a doubled quote in the resourceId field. Treat it as an illustration, not a paste-ready template.

Microsoft's GA post also warns about traces. Foundry traces can contain prompts, model responses, system instructions and tool-call arguments. Recordings and traces are therefore sensitive data in their own right.

The preview additions​

Private websites with VNet injection​

Many valuable workflows run against internal portals, admin consoles and line-of-business apps. With VNet injection, you link your own virtual network and a dedicated delegated subnet to the Playwright Workspace. Cloud browsers can then reach applications inside that network. The post says you can set network boundaries and control browser egress with the networking policies you already apply to the subnet.

This gives the browsers network reachability. It doesn't replace authentication or authorization inside the target application.

Bring your own browser proxy​

Browser proxy lets you route cloud-browser traffic through your own proxy infrastructure. The post describes:

  • configurable proxy pools with domain-based routing
  • priorities, where the highest-priority matching rule decides the proxy
  • bypass patterns
  • proxy credentials held in Azure Key Vault and accessed through managed identities, so they stay out of automation code

Microsoft's sample is an az rest PUT against a proxy-configs endpoint, using preview API version 2026-11-01-preview. It defines a default proxy, an EU route at priority 1 and a corporate route at priority 2. It also bypasses status.company.com and *.blob.core.windows.net. The hostnames and patterns are placeholders. Check syntax and availability against current documentation before relying on them.

Secret management​

Agents need credentials, but the model shouldn't see them. With secret management you keep secrets in your own Azure Key Vault. The Playwright Workspaces remote MCP server reads them server-side, so neither your application code nor the model gets the value. You can bind each secret to a domain allowlist, which stops a malicious site from capturing it. Secret URIs are versionless, so a rotated secret is picked up automatically.

This is a real security boundary, but the work around it remains. You still have to scope credentials narrowly, govern Key Vault access and choose the allowed domains deliberately.

Profile reuse​

Profile reuse is marked preview and "coming soon." It keeps authentication, cookies and browser state between sessions, so long-running or repeated workflows skip repeated logins. Profiles sit in encrypted storage, governed by workspace-level RBAC.

Saved state can hold an authenticated session, so access control and lifecycle management matter. Microsoft's sample references API version 2026-12-01-preview, which is dated after the post itself. That fits the "coming soon" label. Don't assume the endpoint is usable today.

Web Bot Auth and CAPTCHAs​

The section is headed "CAPTCHA management," but Microsoft frames the feature more narrowly. Microsoft says its focus is not getting around a website's security controls. Instead, Web Bot Auth, which the post says is based on an IETF draft, lets agents cryptographically sign HTTP requests. That can help sites verify an agent's identity on public-web research tasks.

There are two routes:

  • self-registration by the customer
  • Microsoft platform registration through selected bot-management providers, named as Akamai, Cloudflare and F5

The sample enables a webBotAuth setting on a Playwright Workspace resource, with signing key Workspace. It uses preview API version 2026-11-10-preview. Signing requests may help a site recognize your agent. It doesn't guarantee the site will accept it, and it doesn't solve CAPTCHAs.

Getting started​

Microsoft's sequence is short:

  1. Create or select a Playwright Workspace. It must be provisioned separately as an Azure resource.
  2. Add the workspace to a toolbox in Microsoft Foundry. Microsoft's documentation says the tool uses the Foundry project managed identity to access the workspace.
  3. Add the toolbox to your agent.
  4. Enable only the capabilities your scenario needs.

If you want private-network access, plan for it early. Microsoft's GA post says private network connection should be added while provisioning the workspace.

The demo and what to make of it​

Microsoft's demo has an agent process 15 invoices. It validates them against rules, flags anomalies in Excel and enters the validated ones into an expense portal. That is a vendor demonstration, not an independent benchmark. It says nothing about accuracy, error rates or cost at scale.

It does show where this tool fits. Many shops have processes that live in a vendor portal or legacy web app with no API. Automating those has usually meant brittle scripts or robotic process automation. A managed browser behind an agent is a different approach. It adapts to page changes, but it adds model-driven behavior to systems that handle money.

Practical takeaways for IT admins​

  • Start narrow. Begin with navigation and inspection tools. Add state-changing tools such as clicking, typing and uploads only when needed.
  • Gate consequential actions. Require human approval for anything that submits or changes records.
  • Protect Live View links. Anyone holding an active session's view link may be able to see or steer the session. Treat the link as sensitive.
  • Use RBAC for recordings and traces. Both can contain sensitive material.
  • Pilot the previews. VNet injection, proxies, secret management and profile reuse can change before they reach GA.
  • Verify before retrying. Browser steps have real side effects. If a step times out, check the application state before re-running it, or you risk duplicate entries.

Microsoft's own August-to-November trail shows how fast this area moves. The GA foundation arrived in October, and the enterprise-networking and credential layer is arriving as preview a month later. The direction is plain: Microsoft wants browser agents to run inside enterprise network, identity and audit boundaries. Whether your risk team accepts that depends on how well you scope each tool, secret and saved session.

 

References

  1. Browser Automation Tool in Microsoft Foundry : Ignite’26 Announcements Microsoft Foundry Blog 2026-11-18T17:00:47+00:00
  2. Automate browser tasks with Foundry agents - Microsoft Foundry learn.microsoft.com
  3. Browser automation tool in Microsoft Foundry is now GA | Microsoft Foundry Blog devblogs.microsoft.com