Browser Agent Cut Flight Search From 30.7 to 18.0 Seconds
Six causes explain why a browser agent took 30.7s versus 17.6s on local Chrome; two config flags alone cut it to 20.7s, and prod now runs 18.0s.
Wire notes
- The agent went from 30.7 seconds on staging (September 27) to 18.0 seconds on prod (October 9).
- Two config flags — service tier and dropping the default proxy — cut Flights time from 30.7 to 20.7 seconds.
- A page-settle detector never fired because Google Flights rewrites a CSS class every frame.
- All fixes made tasks 23–36% faster with all 49 replica runs staying correct.
- A captcha detector burned 32–35% main-thread CPU on an idle Flights page.

By James Calloway
A browser agent that needed 30.7 seconds to find the three cheapest Zurich-to-London flights on Google Flights now runs the same task in 18.0 seconds — and the engineering team behind it says six causes explain 12.8 of the original 13.1-second gap versus a local Chrome run at 17.6 seconds.
The fastest wins were pure configuration. Two flags alone took Flights from 30.7 to 20.7 seconds and example.com from 10.3 to 6.5.
What were the six causes?
- Service tier and proxy. The cloud setup called OpenAI on the standard tier while local runs used priority, cutting mean turn time from 1.20 to 1.02 seconds (p90 from 1.81 to 1.41) — about 1.9 seconds off a 12-turn run. Every run also went through a residential proxy. On October 9, the same harness with a US proxy was 5.3 seconds slower (25.3 vs 20.0, n=8).
- CDP round trips. A laptop reproduction across the country from the cloud browsers paid about 85 ms per CDP call; two steps alone took 8.2 and 8.7 seconds.
- A page-settle detector that never settled. Google Flights rewrites a CSS class on every frame, so the MutationObserver-driven "page is quiet" check ran to its 800 ms cap every time.
- A captcha detector hogging the main thread, blocking it 674–698 ms in every 2-second window and burning 32–35% CPU on an idle page.
- The team's own prompt, adding 1.16x time and 1.70x cost at the same score.
A config surprise: with no explicit model parameters, the harness ran GPT-6 Luna at xhigh reasoning effort. One Flights run took 141 seconds and 39 model calls. The team's advice: set your parameters explicitly.
How did they fix the settle detector?
The fix ignores DOM writes that don't change anything, and counts a style change only if it's the first on its element within 200 ms — a menu opening counts, an animation loop doesn't. Stopping re-probes every 40 ms cut CDP calls on a Flights run by about 30%, with scores, model calls and cost unchanged. Combined, the fixes saved 0.5 to 1.4 seconds per task.
The captcha solver now starts off and turns on only when the agent calls a solve_captcha tool. By October 9 the stall had vanished with the solver on or off; some other browser change in those two weeks fixed it too, and the team never found which one.
Did correctness survive?
On a Cloud replica sitting in the browsers' region, every task got 23 to 36% faster and all 49 runs stayed correct. On prod, October 9, with 8 Flights runs per setup: laptop + local Chrome hit 18.8 seconds at 5/8 correct; Prod Cloud with bu-ultrafast hit 18.0 at 7/8; harness in region plus cloud browser hit 20.0 at 8/8.
One gap remains unexplained: Cloud's model time was 3.4 seconds lower than the same harness calling OpenAI directly (9.6 vs 13.0), with the same turns and the priority tier verified on both. The agents doing the tuning needed watching, too — they edited files mid-benchmark, throwing away five runs, and read judge noise as signal. The team now requires at least 6 runs per task for any quality claim, and the 30 experiment branches from September 26–30 are public, losers included.
The result ships as the bu-ultrafast model setting in Browser Use Cloud — built for short tasks, with bu-fast recommended for long ones.
via browser-use.com (Source)
More from James Calloway
Show full bio
Market editor covering technology and product news in marketing, advertising, marketing technology, media, SEO and search at Marketing Herald.
401 articles
More on the wire
- Proxy Benchmark 2026: Seven Residential Providers Tested on AI Agents
- Lighthouse 13.6 Update Pushes PageSpeed Insights Scores Higher
- OpenAI Pays Cloudflare for Page-Change Hints as 'New SEO' Emerges
- Cloudflare Pilot Uncovered Hundreds of Millions of New Pages for ChatGPT Search
- ChatGPT Referral Crash in GA4 Was a Tracking Bug, Not a Traffic Drop