Skip to content

Open-source Python library for AI browser agents, plus Browser Use Cloud: stealth browsers and a hosted agent API.

Record post fallback eval gate

Magnus Müller5d716ea439b3
diff --git a/RUST_AGENT_INTEGRATION_FEATURES.md b/RUST_AGENT_INTEGRATION_FEATURES.md
index 6a80c3146..7d4c89040 100644
--- a/RUST_AGENT_INTEGRATION_FEATURES.md
+++ b/RUST_AGENT_INTEGRATION_FEATURES.md
@@ -1480,6 +1480,11 @@ Terminal core branch: `magnus/browser-use-rust-main-integration` at terminal mai
- All five Browser Use cloud sessions were created promptly and all five tasks entered Rust-core agent execution. Agent-stage completions before the 10-minute stop were task 2 at 162.06s, task 3 at 168.49s, task 1 at 225.02s, and task 5 at 227.80s.
- Saved results before the stop: task 2 score `0.0`, task 3 score `0.9`, and task 1 score `0.0`; task 1's `0.0` came from Agent SDK judge max-turn exhaustion, which feature 246 addresses.
- The gate showed the 7-second threshold improved the agent-stage tail for several tasks compared with the prior 403s/470s/488s long tasks, but did not justify scaling to 50 because task 4 and serialized Agent SDK judging still exceeded the 10-minute gate.
+- 2026-06-04 post-judge-fallback five-task real_v8 Agent SDK gate:
+ - Run `kh74xjt5jwnhbj5kz11k7xc29s881784` used `BU_RUNTIME=brust`, `--browser browser-use-cloud`, `BROWSER_USE_CLOUD_PROXY_COUNTRY_CODE=us`, `BU_BROWSER_SCRIPT_INITIAL_WAIT_MS=7000`, `--judge-type agent_sdk`, `--model claude-sonnet-4-6`, `--eval-model claude-sonnet-4-6`, `--parallel-runs 5`, and `--max-steps 100`.
+ - All five Browser Use cloud sessions were created promptly and all five tasks entered Rust-core agent execution. Saved results before the 10-minute stop: task 1 score `0.0` in 89.98s agent time/125.52s total pipeline, task 2 score `0.0` in 161.80s/213.03s, and task 3 score `1.0` in 282.02s/343.87s.
+ - Tasks 4 and 5 were still running when the 10-minute limit was reached; their remaining Browser Use cloud sessions were manually stopped. No Agent SDK judge exception or max-turn fallback was observed in the saved results.
+ - This gate still did not justify scaling to 50 because the sample quality was below target and the long-tail tasks did not complete within the 10-minute iteration budget.
- 2026-06-03 three-task real_v8 cloud smoke through the Python `Agent` API and Rust terminal core launched and returned for real_v8 `18`, `14`, and `20` with `DEFAULT_LLM=openai_gpt_4_1_mini`, `BROWSER_USE_RUST_BROWSER_MODE=cloud`, and the local debug `browser-use-terminal`, but it is not a clean benchmark pass:
- real_v8 `18`: returned `successful=true`, but final output was `[empty]` after failing to verify AND Digital leadership information and included generated browser-script JavaScript quoting errors.
- real_v8 `14`: returned `successful=true`, but final output said ycombinator.com was blocked/unresponsive and did not extract the requested Winter 2025 B2B company data.