AgentArmorHQ
Sign in
  • AgentArmor teamTeam3d agoin s/browser-use-browser-useNeeds helpSolved

    browser-use agent stuck at Step 1 with Chromium on about:blank page

    Using browser-use 0.1.40 with Azure OpenAI GPT-4o on Windows 11, my agent opens Chromium on an about:blank page and then loops forever, printing the same 'Step 1' log line repeatedly. It only continues if I manually type something into the Chromium address bar and press Enter. Is there a known fix or configuration change for this stuck-at-startup behavior?

    Originally asked on GitHub: https://github.com/browser-use/browser-use/issues/839

    6 answers

    • Best answerAgentArmor teamTeam3d ago

      Closed as completed, but the thread does not confirm one universal fix. The looping INFO [agent] 📍 Step 1 message with Chromium left on an about:blank tab was reproduced by several people: Azure OpenAI gpt-4o on Windows 11 (browser-use 0.1.40), OpenAI gpt-4o on macOS 15, DeepSeek, and a locally hosted DeepSeek R1 via Ollama. No single root cause was confirmed, but two things helped some reporters:

      1. Check your LLM API credits. One macOS user using OpenAI saw the identical loop and found they were out of OpenAI credits; after topping up, the same code worked without changes. Look at your provider dashboard before debugging further.
      2. Make the task start with a full, explicit URL. One Azure OpenAI user (gpt-4o, browser-use 0.1.40) tested task phrasings in order: a plain task like Compare the price of gpt-4o and DeepSeek-V3 looped forever, and so did Start from www.google.com and Go to www.google.com (no scheme). Only tasks starting with the full URL worked, e.g. Go to http://www.google.com/, then compare the price of gpt-4o and DeepSeek-V3. Another user tried that exact full-URL phrasing and still hit the loop, so this is model/provider dependent, not a universal fix.

      Also worth trying: run from the current main branch. A maintainer said all the failing phrasings worked there, although another user who cloned main on macOS still reproduced the loop with OpenAI gpt-4o, so that was not conclusively confirmed either.

      General advice: when the agent is stuck repeating Step 1, first verify the model call itself is succeeding (credits, endpoint, model availability), then try rewriting the task to begin with an explicit, full URL. The Azure OpenAI setup in the report loaded AZURE_OPENAI_API_KEY from a .env file; keep that file out of version control.

      Did this work for you?
    • AgentArmor teamTeam3d ago

      What a GitHub user tried on GitHub (Didn't work)

      Reporter ran the basic agent sample on Windows 11 with Azure OpenAI gpt-4o and browser-use 0.1.40, using the plain task 'Compare the price of gpt-4o and DeepSeek-V3' with no URL. Chromium opened on about:blank and the agent printed the Step 1 line forever.

      Source: https://github.com/browser-use/browser-use/issues/839 (opens external site)

      Did this work for you?
    • AgentArmor teamTeam3d ago

      What @APantelides tried on GitHub (Worked)

      A macOS user saw the same loop with OpenAI and their local Chrome install, then discovered they were out of OpenAI credits. After adding credits the same code worked.

      Source: https://github.com/browser-use/browser-use/issues/839#issuecomment-2677250056 (opens external site)

      Did this work for you?
    • AgentArmor teamTeam3d ago

      What @MrMarsMic tried on GitHub (Worked)

      An Azure OpenAI user tested progressively more explicit task phrasings. Plain text, 'Start from www.google.com', and 'Go to www.google.com' all looped; 'Go to http://www.google.com/, then compare the price of gpt-4o and DeepSeek-V3.' completed successfully.

      Source: https://github.com/browser-use/browser-use/issues/839#issuecomment-2678042270 (opens external site)

      Did this work for you?
    • AgentArmor teamTeam3d ago

      What @andreadsw tried on GitHub (Didn't work)

      A macOS 15 user cloned the main branch, used OpenAI gpt-4o, and still got the same stuck-on-about:blank behavior, contradicting the maintainer's claim that main branch worked.

      Source: https://github.com/browser-use/browser-use/issues/839#issuecomment-2681933658 (opens external site)

      Did this work for you?
    • AgentArmor teamTeam3d ago

      What @fromfromto tried on GitHub (Didn't work)

      Another user tried the exact phrasing 'Go to http://www.google.com/, then compare the price of gpt-4o and DeepSeek-V3.' that worked for comment 3, and it still failed with the same Step 1 loop.

      Source: https://github.com/browser-use/browser-use/issues/839#issuecomment-2684195678 (opens external site)

      Did this work for you?

    Add your answer or comment

See Browser Use, its grade and setup notesCheck this tool before you install