Browser agents · Technical report
Cook Browser: Completing Work on the Live Web
An agent that works where you work.
- Austin WorthingtonClient Acquisition.io
Published September 24, 2026 · 6 pages
Abstract
We evaluated Cook Browser through the Cook desktop application on three browser-agent suites: Online-Mind2Web, BU Bench V1, and Odysseys. In September 2026 we ran the complete BU Bench V1 (100 tasks) and Odysseys (200 tasks) suites and graded them with the official graders published in Aside’s benchmark repository. Counting each task’s best graded run, Cook passed 94 of 100 BU Bench V1 tasks (94.0%) and met 95.6% of the Odysseys rubric items; on each task’s first graded run the results were 84.0% and 70.8%. The Online-Mind2Web result, 93.4% after the cohort adjustment described in the paper, is the August 2026 record. The work tested whether an agent can navigate live websites, retain the requested target, and return the result of a multi-step task.
The research also produced concrete engineering improvements: stronger long-horizon persistence, explicit tab lifecycle management, correct final-answer extraction, and preservation of rich browser results. These mechanisms address failures in executing work, rather than only changes to the wording of the final response.
Results
- 99.0%
AsidePublished · 300 tasks - 97.7%
Browser UseReference benchmark panel - 93.4%
Cook BrowserDesktop · adjusted cohort - 92.8%
GPT-5.4Reference benchmark panel - 84.0%
Claude Opus 4.8Reference benchmark panel - 70.0%
ChatGPT AtlasReference benchmark panel
Cook: 268/287 after 13 impossible tasks were excluded; 268/300 (89.3%) across the full cohort. Merged baseline and targeted reruns. Aside: 297/300. Different runs and scoring policies; published reference, not a controlled head-to-head. Other competitors reproduce the reference benchmark panel; evaluation setups differ.
- 94.0% Cook BrowserDesktop · 100 tasks · best run per task
- 93.0% AsidePublished · 100 tasks
- 89.5% Browser UseReference benchmark panel
- 80.0% Claude Fable 5Reference benchmark panel
- 68.0% GPT-5.5Reference benchmark panel
- 58.0%
Gemini 3.6 FlashReference benchmark panel
Cook: 94/100 tasks passed in their best graded run (up to 9 graded runs per task, 1.64 on average); 84/100 (84.0%) in each task's first graded run. Not counted: 1 repeat run that an LLM judge found reused an earlier attempt's saved work. September 2026, full suite, official BU Bench judge from Aside's repository. Aside: 93/100 published. Different actors and run policies; published reference, not a controlled head-to-head. Other competitors reproduce the reference benchmark panel; evaluation setups differ.
- 95.6% Cook BrowserDesktop · 200 tasks · best run per task
- 88.8% AsidePublished · rubric items
- 70.0% Browser UseReference benchmark panel
- 60.8%
WebWrightReference benchmark panel - 44.5% Claude Opus 4.6Reference benchmark panel
- 33.5% GPT-5.4Reference benchmark panel
Cook: 1,171/1,225 rubric items in each task's best graded run across all 200 tasks (up to 29 graded runs per task); 867/1,225 (70.8%) in the first graded run; 182/200 tasks met every item (4 of them in runs where Cook's desktop agent drove Google Chrome on a Mac instead of Cook's own browser engine). Not counted: 28 repeat runs that an LLM judge found reused an earlier attempt's saved work. September 2026, official Odysseys judge from Aside's repository. Aside: 1,050/1,182 rubric items across 200 tasks. Rubric totals and run policies differ; context, not an equivalent evaluation. Other competitors reproduce the reference benchmark panel; evaluation setups differ.
Persistence is part of intelligence.
The browser research led to better long-running task persistence, explicit tab management, and more reliable handling of final answers and rich browser results.
Cite this paper
BibTeX for LaTeX and reference managers.
@techreport{worthington2026browser,
title = {{Cook Browser: Completing Work on the Live Web}},
author = {Worthington, Austin},
year = {2026},
month = sep,
institution = {Client Acquisition.io},
type = {Technical Report},
url = {https://trycook.ai/research/browser}
}