Benchmark report
Task verified
5m 5s elapsed
The saved attempt met the documented success criterion.
96/ 100
Agent-Ready- SInstall
- SAuth
- SExecute
- ADocs
- SCurrent
A few dead doc URLs (guessed stop-browser-session.md 404, api-reference/v4 list page empty, auth.md proxy refusal) cost three extra fetches before the right api-v4 pages.
Where the time went
5:06 agent time- Starting0:01
- Research0:47
- Attempt3:05
- Verify1:13
Judge reason
Saved code artifacts/browser_session_lifecycle.py uses only stdlib urllib against https://api.browser-use.com/api/v4 with the key sent solely in the X-Browser-Use-API-Key header. Transcript entry 42/45 shows the script executed live (EXIT=0) producing: POST /browsers body {"timeout":5} -> HTTP 201 with id f7038328-3912-4f83-bbdb-b3f2e6ee1d42, startedAt 05:19:11.198892Z and timeoutAt 05:24:11.198892Z (exactly a 5-minute window, confirming timeout is minutes); GET /browsers/{id} -> 200 status "active"; GET /browsers?pageSize=100&pageNumber=1&filterBy=active -> 200 with the same id in items (03_list_membership.json present_in_list true); PATCH /browsers/{id} {"action":"stop"} -> 200 status "stopped", finishedAt 05:19:11.814665Z; final GET -> 200 status "stopped". Termination and the final re-fetch are inside the finally: clause of main() (lines 123-145). No Playwright/CDP/websocket import or connection; cdpUrl/liveUrl redacted. Console log values match the saved JSON responses exactly, and an in-run grep -rqF of the key over artifacts and outputs returned "no-key-leak: clean" (my own scan found no bu_/sk- pattern). All four operations are documented (DOCS_NOTES.md cites the create/get/list/update api-v4 pages fetched in the transcript), so no Not-supported report was needed. Responses are live and non-mocked (real UUID, cost fields, monotonic timestamps).
Attempt record
- 2:04Answered from saved key: Browser Use Cloud API keyno wait
- —First success verified
Attempt artifacts saved.