Share via

Azure AI Foundry Agent Service – Browser Automation tool returns empty output ("") with status OK

Edmond Varga 11 Reputation points
2026-02-17T14:33:24.3766667+00:00

I’m having an issue with the Browser Automation (preview) tool in Azure AI Foundry Agent Service.

Scenario

  • Region: eastus2
  • Agent model: gpt-4.1/gpt-5
  • Tool: Browser Automation (preview)
  • Playwright workspace: connected per the official guide, permissions verified

Problem

When the agent calls Browser Automation, the run completes but the tool returns an empty string. There’s no error and the span duration is 0ms.

Reproduction

  1. In Foundry Agent Playground, add Browser Automation tool (connected to a Playwright workspace)
  2. Prompt: “Open https://example.com and return the page title”
  3. In Traces, the tool output is empty

Trace snippet

{
  "name": "remote_function_call_output: browser_automation",
  "attributes": {
    "duration": 0,
    "status": "completed",
    "name": "browser_automation",
    "output": ""
  },
  "status": { "status_code": "OK" }
}

What I tried

  • Followed the setup/connect Playwright workspace guide
  • Double-checked app registration permissions
  • Tried using different models
  • Read troubleshooting docs (no error shown)

Additional detail

If I connect to the same Playwright workspace directly (simple JS test + Entra ID), it returns DOM/content as expected. So the workspace works; the issue seems specific to the Agent tool integration.

Question

What causes Browser Automation in Foundry Agent Service to return output="" with status OK/duration=0, and what should I verify (region/feature flags/RBAC/Playwright workspace binding/permissions) to make it actually execute and return results?

Azure AI services
Azure AI services

A group of Azure services, SDKs, and APIs designed to make apps more intelligent, engaging, and discoverable.

1 answer

Sort by: Most helpful
Most helpful Newest Oldest
  1. SRILAKSHMI C 15,030 Reputation points Microsoft External Staff Moderator
    2026-03-06T11:39:55.04+00:00

    Hello Edmond Varga,

    Welcome to Microsoft Q&A and Thank you for reaching out.

    From the behavior you described (status = OK, duration = 0 ms, and output = ""), it typically indicates that the tool invocation completed at the agent level, but the browser action itself was never executed. This usually points to a configuration, permission, or preview feature issue rather than a failure in the Playwright workspace itself.

    Since you mentioned that the Playwright workspace works correctly when tested directly, the issue is most likely within the integration between the Agent service and the workspace in Azure AI Foundry.

    Below are a few areas worth verifying:

    1. Region and preview feature availability

    The Browser Automation tool is currently in preview, and availability may vary by region or subscription feature flags. Even though your agent is running in East US 2, the preview capability must also be enabled for your environment. If the feature is not fully enabled, the tool call can complete immediately with 0 ms duration and empty output, which matches the behavior you’re seeing.

    2. Playwright workspace connection

    Even if the workspace works independently, verify that the Agent tool connection is configured correctly:

    Confirm the Target URI for the Playwright workspace is correct and begins with wss://.

    Ensure the Access Token or authentication method is valid.

    Verify the workspace connection shows Connected within the Agent tool configuration.

    Ensure the same tenant and subscription are used for both the Agent project and the Playwright workspace.

    3. RBAC and project identity permissions

    The Agent project identity must have permission to access the Playwright workspace.

    Check that the identity used by the agent has:

    Contributor role or equivalent custom role

    Access to perform browser execution operations on the workspace.

    If the identity lacks permissions, the tool may return an empty output without throwing an explicit error.

    4. Model and region compatibility

    Since you mentioned using GPT-4.1 and GPT-5, verify that the model + region combination is supported for agent tools in your environment. Unsupported combinations can sometimes cause tool steps to be skipped silently.

    5. Prompt and tool invocation

    Ensure the agent is actually invoking the browser tool.

    If needed, you can also explicitly configure the tool_choice parameter to force the agent to use the Browser Automation tool.

    6. Trace inspection

    In the Agent traces, check if there is a tool call request span before the remote_function_call_output span.

    Specifically verify:

    Whether the navigation action (URL, DOM query, etc.) is present in the tool arguments.

    If the tool request span is missing or empty, which could indicate the planner skipped execution.

    7. Network configuration

    If your environment uses VNET integration, private endpoints, or restricted outbound networking, verify that the agent service can reach the Playwright workspace endpoint. Network restrictions can sometimes prevent execution even though the tool call appears successful.

    Troubleshooting steps

    You may also want to try:

    Re-running the Browser Automation setup step-by-step following the official documentation.

    Enabling additional logging or network tracing during the agent run.

    Testing with a very simple navigation command to isolate whether the issue is related to the task complexity or the integration itself.

    I Hope this helps. Do let me know if you have any further queries.

    Thank you!

    0 comments

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.