← All agent and harness briefs
Agent runtimeagenticopsresearchdocumentedunattended capable

Operational brief

Browser Use

Python browser agent, CLI controls for your existing agent, and a separate fully hosted agent API.

Official Browser Use image from browser-use.com

Choose it when

You want a customizable Python browser agent, controls for an existing agent, or hosted V4, and can choose the model, browser location and host isolation appropriate to the task.

Skip it when

You need guaranteed approval before every external action, host isolation from a remote browser alone, or local-only defaults without configuring the model and browser. Browser infrastructure alone doesn't supply the agent logic.

01 / Classification

Autonomy evidence

This is our classification, not a vendor badge. The record says what the product normally does. The gaps stay visible.

How we classify it
Tool use + multi-step plans, still human-gated per session.
Runs unattended
Yes — that capability is documented in this brief.
Trigger
Submit a hosted V4 prompt, run the Python agent in your application, or have an existing agent call the CLI.
State
For V4, a session holds conversation, a workspace holds files and a profile holds browser login state; these IDs aren't interchangeable.
Tools
The 1.0 Python agent writes Python against CDP in a persistent REPL running as the host user. The CLI gives an existing agent browser controls; hosted V4 runs both agent and browser remotely. Browser-only infrastructure doesn't supply the agent loop.
Human control
Design hosted human checkpoints into the task: wait for the run to stop before browser takeover, then continue. Treat the live-view URL as a credential. This isn't automatic approval before every action; the task owner must reconcile important writes.

02 / Operations

Control reality

Useful agents eventually touch something real. This is the operational picture we have, not a promise hidden in a sales page.

Where it runs
The Python agent runs on your host but defaults to a cloud browser; an explicit local browser is available. The CLI serves an existing agent, while hosted V4 runs both agent and browser remotely. A remote browser doesn't isolate host Python.
Data boundary
The Python library defaults to the Browser Use model gateway and cloud browser; local execution doesn't keep browser context off providers. Explicit Profile Sync copies selected local Chrome cookies to a cloud profile. Screenshots aren't secret-redacted, and the documented domain guard isn't a network sandbox.
Identity & permissions
Existing local Chrome login state, a copied local profile and a cloud profile are distinct. The default Browser Use model gateway needs its own access; alternative model credentials can be configured. Treat live-view URLs as credentials granting browser control.
What it can write
browser via CDP · host-user Python file, process and network operations · workspace files and downloads
Surfaces
Python library · CLI · Browser Use cloud
Export / lock-in
Library 1.0 removes the storage_state option; result.messages exposes the transcript and result.downloads lists run files. Hosted V4 workspaces hold files separately from conversation sessions and browser profiles. Full portable export and restore remain unverified.
Operational limits
Open source refers to the library, not free hosted service. Library 1.0 is a breaking generation: its Pydantic AI loop writes Python in a persistent REPL running as the host user, even with a remote browser; that isn't a sandbox. A cloud browser alone isn't the hosted agent. Human checkpoints must be designed into tasks, and a client timeout doesn't cancel a hosted run. The earlier OpenAI-driver and Anthropic SDK caveats are historical, not current compatibility findings. Documented, not runtime-tested.
Watch for
Python library 1.0.6 defaults to bu-4.0 on the Browser Use gateway and a cloud browser; explicit local browser and model alternatives exist. Model-written Python runs as your host user, even with a remote browser. Separately, Cloud API V4 US projects created on or after November 16, 2026 default to bu-4.0 when model is omitted and High reasoning when modelParams is omitted; Bedrock-only projects retain their compatible provider default. Older US projects keep their documented default and EU defaults are unchanged; explicit values override the defaults. Don't apply that project rule to every installation. A client wait timeout doesn't cancel the hosted run; reconcile run status and verify important external writes before retrying. A model-reported done or success isn't independent proof that an external action succeeded. Live-view URLs grant browser control.

03 / Receipts

Evidence ledger

Here is what this brief is built on. “Documented” means we found a source. “Tested” means this desk actually ran it. Those are different things.

Evidence status
documented
Source set
13 primary sources
Evidence review
2026-11-17

Corrections

  • 2026-10-13 · Replaced the June CLI/CDP desk-report timestamp with library 0.13.11 publication. Separated hosted V4 agent, existing-agent CLI and Python library from browser-only infrastructure. Corrected unattended:false to documented optional hosted continuation, not autonomous reliability. Retained historical payloads; added SDK availability and checkpoint/data-boundary limits.
  • 2026-10-16 · Refreshed the Python library release to 0.13.13, not Cloud API V4. The release supersedes 0.13.12, whose PyPI publication was cancelled before approval. Added OpenAI driver prerequisites and the current integration guide's optional application-host Bash boundary; a remote browser isn't a host-tool sandbox. The prior 0.13.11 Anthropic SDK warning is historical, not a current compatibility finding. Documented, not runtime-tested.
  • 2026-11-17 · Updated the Python library to the breaking 1.0 generation, release 1.0.6 published 2026-11-16T00:51:24Z in the retained release API. Replaced the earlier generation's current-driver assumptions with the documented host Python boundary; Cloud V4 remains a separate product. No installation or runtime test.

04 / Change log

Related coverage

The brief tells you what the product is. Payloads track the changes that may alter the decision.