Skip to content

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

Record summary cap eval gate

Magnus Müller245e99ae82ec
diff --git a/RUST_AGENT_INTEGRATION_FEATURES.md b/RUST_AGENT_INTEGRATION_FEATURES.md
index be6e8410b..89e603039 100644
--- a/RUST_AGENT_INTEGRATION_FEATURES.md
+++ b/RUST_AGENT_INTEGRATION_FEATURES.md
@@ -1561,6 +1561,14 @@ Terminal core branch: `magnus/browser-use-rust-main-integration` at terminal mai
## Not Verified Yet
+- 2026-06-04 post-summary-first-browser-script five-task real_v8 Agent SDK gate:
+ - Run `kh77e9y1cjeyygdnvex85qwcrs880hr0` used `BU_RUNTIME=brust`, terminal commit `297930a`, `--browser browser-use-cloud`, `BROWSER_USE_CLOUD_PROXY_COUNTRY_CODE=us`, `BU_BROWSER_SCRIPT_INITIAL_WAIT_MS=7000`, `AGENT_SDK_JUDGE_TIMEOUT_SECONDS=25`, `AGENT_SDK_JUDGE_FALLBACK_ATTEMPTS=3`, fallback timeout default floor `45`, `AGENT_SDK_JUDGE_CONCURRENCY=16`, `LLM_BROWSER_CAPTURE_FPS=0`, `BU_DISABLE_FALLBACK_CAPTURE_GIF=1`, `BU_BROWSER_SCRIPT_SESSION_OUTPUTS=1`, `--use-vision false`, `--judge-type agent_sdk`, `--model claude-sonnet-4-6`, `--eval-model claude-sonnet-4-6`, `--parallel-runs 5`, and `--max-steps 100` for real_v8 tasks 6-10.
+ - Browser Use cloud startup was healthy: all five sessions were created around 2026-06-04T10:54:10Z, all five tasks entered Rust-core agent execution, and no CAPTCHA/local-browser misconfiguration appeared.
+ - Saved results before the 10-minute cap: task 10 score `1.0` in 229.10s agent time/289.97s total pipeline, task 9 score `0.75` in 312.69s/391.38s, and task 8 score `0.8` in 456.07s/532.48s.
+ - The prior fatal Agent SDK `success` exception did not recur. Saved tasks used the bounded judge fallback after 25-second tool-judge timeouts and saved scores normally.
+ - Compared with the previous 4 KiB gate, task 10 improved materially in agent time (`383.78s` to `229.10s`) while task 9 regressed (`222.07s` to `312.69s`) and task 8 completed/saved instead of remaining unsaved.
+ - Tasks 6 and 7 were still in agent-stage execution when the cap fired. All five Browser Use cloud sessions were explicitly stopped after the cap, and no eval/terminal subprocesses remained.
+ - This gate does not justify scaling to 50: saved-task mean score was `0.85`, but missing-task-as-zero mean score was `0.51` with only 3/5 tasks saved within the iteration budget. The remaining bottleneck is still long-tail document/research execution and cumulative prompt growth, not cloud browser startup or judge exception handling.
- 2026-06-04 post-4k-browser-script-cap five-task real_v8 Agent SDK gate:
- Run `kh75ejqt6j76rc4k0xecf1rmq9881h9t` used `BU_RUNTIME=brust`, `--browser browser-use-cloud`, `BROWSER_USE_CLOUD_PROXY_COUNTRY_CODE=us`, `BU_BROWSER_SCRIPT_INITIAL_WAIT_MS=7000`, `AGENT_SDK_JUDGE_TIMEOUT_SECONDS=25`, `AGENT_SDK_JUDGE_FALLBACK_ATTEMPTS=3`, fallback timeout default floor `45`, `AGENT_SDK_JUDGE_CONCURRENCY=16`, `LLM_BROWSER_CAPTURE_FPS=0`, `BU_DISABLE_FALLBACK_CAPTURE_GIF=1`, `BU_BROWSER_SCRIPT_SESSION_OUTPUTS=1`, `--use-vision false`, `--judge-type agent_sdk`, `--model claude-sonnet-4-6`, `--eval-model claude-sonnet-4-6`, `--parallel-runs 5`, and `--max-steps 100` for real_v8 tasks 6-10.
- Browser Use cloud startup remained healthy: all five sessions were created around 2026-06-04T10:34:51Z, and all five tasks entered Rust-core agent execution promptly.