Vibeleaderboard
← All Intel
Intel / repo

browser-use 1.0.8 — browser-use 1.0.8

Source
github.com
Author
gregpr07Top Viber
Date

Context

Browser Use, an open-source Python library in which an AI model completes tasks in a web browser by writing code, released 1.0.8 as a set of fixes, and the release notes show that most of them address runs that ended without an answer or cost more than they needed to. The change with the widest reach is to the judge. After each run that took a step, Browser Use sends one more request to the same model asking it to grade the run against the task, and the answer records whether the run succeeded, why it failed, whether it met a , bot block or CAPTCHA, and a few gotchas, short lessons about the sites for the next agent. Until now that request sent the run's trace as a new , which the provider billed in full. According to the pull request, 1.0.8 asks the judging question as one more turn of the run instead, with the run's own messages, instructions and tool definitions first, so the provider can read that unchanged opening from its prompt cache, a store of recently processed prompt text, and bill it as cached input. If the model answers with a or without the expected fields, Browser Use falls back to the old separate request. The notes give no figure for the saving. The same pull request also stops the judge from mistaking an OpenAI rate-limit error, whose body names an image rate-limit header, for a model rejecting screenshots, a mistake that had left runs unjudged.

Three other fixes target runs that failed partway or ended without an answer. When a run reaches its step or time limit, Browser Use gives the model a final turn in which it must call a tool that delivers its answer. DeepSeek and Qwen models in thinking mode, along with gpt-oss and models served through vLLM, reject that forced choice, and the pull request says the run then ended in an error with no result; 1.0.8 asks again and leaves the choice to the model. Browser(cdp_url=...) connects the agent to a browser that is already running, through Chrome DevTools Protocol, the interface Chrome exposes so other programs can control it. The agent used to open a blank tab of its own, so an agent sent to a page its caller had prepared, for example with Playwright, saw about:blank and gave up (pull request). It now takes over the open page when there is exactly one; when there are several, the agent can pick one from its list of tabs. Finally, Browser Use's gateway, its hosted model service billed to a Browser Use key, holds enough credit for a call's worst case, the whole input plus the maximum output, before running it. With output capped at 32,768 tokens, small balances were refused calls they could have paid for. Gateway calls now cap output at 16,384 tokens, which the pull request says agent turns stay far below, while other providers keep their settings.

The last two changes concern keys and secrets. Since the 1.0 rewrite made a single Browser Use key the default way to pay for both the model and a cloud browser, a run started without a key has to get one first. On a desktop, Browser Use already opened a login page and continued after approval. Over SSH or on a Linux machine without a display, the browser that approves cannot reach the machine running the agent, so 1.0.8 prints a link and a code to approve on any device and then continues; according to the pull request, CI runs and notebooks keep the missing-key error, as do containers unless a setting allows the code. Separately, bu.type, the helper that types into page fields, no longer writes a password 's contents into its action line. The pull request says that line had put the plaintext password into the run's trace, logs and judge prompt; it now shows only the number of characters.

Key takeaways · AI-distilled
  • In browser-use 1.0.8 the judge runs as one more turn of the same run, so the provider reads the run from its cache instead of billing the trace again, and the verdict keeps the same fields, gotchas included.
  • When DeepSeek or Qwen in thinking mode refuses the forced final call at the step limit, browser-use 1.0.8 asks again without forcing, so the run still returns its answer.
  • Browser(cdp_url=...) now attaches to the one page already open in the connected browser rather than opening a new blank tab.
  • Browser Use gateway calls now reserve credit for 16,384 output tokens instead of 32,768, so accounts with small balances are no longer refused mid-run.
Terms in this piece · Glossary
  • prompt injection — An attack that hides instructions in content an AI will read — a webpage, email, or document — tricking it into following the attacker instead of the user.
  • Prompt — Information or instructions supplied to a model; in a chat, the request you type is often called your prompt.
  • Tool call — Text the model generates in an agreed format that names a tool and gives its inputs.
  • Field — One named piece of a record, such as name or approved_on; a record is one entity's set of values, one per field.
More from gregpr07
Recommended reads
Comments

Checking sign-in…

Loading comments…