WebAgentlab (@webagentlab) header
WebAgentlab (@webagentlab) profile picture

WebAgentLab is building an open-source community focused on Web Agent and the broader GUI Agent field.

join to contribute ๐Ÿ‘‰
Joined November 2024
Photo by WebAgentlab (@webagentlab)Photo by WebAgentlab (@webagentlab)Photo by WebAgentlab (@webagentlab)Photo by WebAgentlab (@webagentlab)Photo by WebAgentlab (@webagentlab)Photo by WebAgentlab (@webagentlab)Photo by WebAgentlab (@webagentlab)Photo by WebAgentlab (@webagentlab)Photo by WebAgentlab (@webagentlab)Photo by WebAgentlab (@webagentlab)
๐Ÿš€ A must-join knowledge hub for GUI Agent builders Building GUI Agents today is noisy. Papers explode, products iterate weekly, and real signal is hard to track. Thatโ€™s exactly why WebAgentLab exists. An open-source community focused on GUI Agents, with 5,000+ members across academia and industry. More than a knowledge base โ€” itโ€™s a collaborative brain for the GUI Agent era: ๐Ÿ”น Curated industry briefs (daily โ†’ weekly โ†’ monthly) ๐Ÿ”น Structured paper database to spot trends fast ๐Ÿ”น Top conference guides & global event radar ๐Ÿ”น GUI Agent product landscape & hands-on evaluations ๐Ÿ”น Open-source collaboration toward industry standards ๐Ÿ”น High-signal job matching inside the core circle If youโ€™re building, researching, or betting on GUI Agents this is where the signal lives, not the noise. Feishu knowledge base webagentlab.feishu.cn/wiki/Zโ€ฆ follow us on Xiaohongshu (RED)
1
1
11
1,121
WebAgentlab retweeted
Today, we're releasing SolidWorks-Bench: can AI models do real mechanical engineering tasks in SolidWorks? Models get close on individual parts but struggle to assemble them. No model completes a single machine. Fable 5.2 leads with a 57% mean reward. Full report below
48
23
13
464
54,592
WebAgentlab retweeted
I asked 56 people in AI which benchmarks they'd trust most to indicate if an LLM is frontier: 25 votes: Terminal-Bench 4 (@alexgshaw, @andykonwinski, @Mike_A_Merrill) 20 votes: Terminal-Bench Science (@StevenDillmann, @lschmidt3, @sanmikoyejo) 13 votes: FrontierCode (@penlume, @ybenpan, @denizbirlikci, @silasalberti) 10 votes: FrontierSWE (@evan_j_chu, @_rajanagarwal, @MatternJustus, @ardnoKhtnayhsiR, @sanket__mhatre, @akshitkr) 8 votes: Humanity's Last Exam (@longphann_, @hendrycks, Alice Gatti, @ziwen_h, @natliml) 7 votes: Vending-Bench (@axelbacklund, @lukaspet) 5 votes each: - AutomationBench (@danielwshepard, @SalimansRobin, @TweetAnnaMarie) - DeepSWE (@winkey_h, @charleyslee, @LeonardMainnet, @serenaa_ge) - PostTrainBench (@full__rank, @hrdkbhatnagar, @AmyPrb, @shiraeis, @karinanguyen, @MatthiasBethge, @maksym_andr) The most common comment was that people don't really trust benchmarks at all. 56 people voted. People couldn't pick a bench their company worked on, an index, or a human-voted board. Disclosure: on this list, I'm a small investor in Abundant, Cognition, Emulated, Fleet, Hillclimb, Latch, and Snorkel. I didn't vote. Counts are net: a few benchmarks received negative votes. Continued:
47
57
36
464
87,582
WebAgentlab retweeted
1/ We measured with Scale AI how well computer-use agents work on each OS, and where they fail today. On Omarchy, Cua Driver lets an agent work while you keep typing: 97% of trials left your keyboard, mouse and windows alone, vs 83% on macOS, 41% on Windows and 40% on Ubuntu.
16
20
4
223
15,649
WebAgentlab retweeted
๐Ÿš€ Excited to unveil PrismaCoder: our latest push in Multimodal Code Intelligence! We built the largest-ever visual-coding dataset (2M!) and trained open-source models that rival Claude Opus 4.6 & Gemini 3.1 Pro. ๐Ÿ”ฅ All data, models & code are open! ๐Ÿ‘‡ ๐Ÿ  Homepage: qiushisun.github.io/PrismaCoโ€ฆ ๐Ÿค— HF: huggingface.co/papers/2610.2โ€ฆ
3
11
1
35
1,432
WebAgentlab retweeted
Microsoft Research's WildAWM A scalable pipeline that synthesizes executable computer-use agent environments and tasks using only open-source models, yielding 341 long-horizon env-task pairs and Qwen agents that generalize to 8 unseen benchmarks.
4
8
1
59
3,815
CUA-Lite: An Open Platform for Computer Use Agents Zhanhui Zhou, Weichen Zhang, Haoran Liu, Lingjie Chen, Tianneng Shi, Kevin Lin, Zhengyuan Yang, Lijuan Wang, Dawn Song arxiv.org/abs/2610.26071 [๐šŒ๐šœ.๐™ฐ๐™ธ]
1
1
256
WebAgentlab retweeted
People long believed in human-like computer use. But software (even terminals) is built for humans, not AI agents. Why not let agents build their own? Meet ๐—”๐—ด๐—ฒ๐—ป๐˜ ๐—ฆ๐Ÿฐ, an agent scientist evolving its own interfaces, letting an open 27B model beats GPT-5.5 & Opus 4.8. With its evolved interfaces, the same model improves ๐Ÿฏ.๐Ÿดร— and costs <๐Ÿฑ% as much per task as either frontier model. Agents today read screenshots and act through click, scroll, type, maybe a shell. So an agent fills a spreadsheet cell by cell when one line of code would do, or prints two thousand lines of a file to find one function. Every new app or model release means a human reads trajectories, diagnoses failures, and hand-writes new primitives. Skill-distillation methods can reorder what the interface allows, but they canโ€™t extend it. Agents compensate with more test-time compute instead of a better interface. ๐— ๐—ผ๐—ฑ๐—ฒ๐—น๐˜€ ๐—ถ๐—บ๐—ฝ๐—ฟ๐—ผ๐˜ƒ๐—ฒ; ๐—ถ๐—ป๐˜๐—ฒ๐—ฟ๐—ณ๐—ฎ๐—ฐ๐—ฒ๐˜€ ๐˜€๐˜๐—ฎ๐˜† ๐—ณ๐—ฟ๐—ผ๐˜‡๐—ฒ๐—ป. โ€œWe want AI agents that can discover like we can, not which contain what we have discovered.โ€ โ€” Rich Sutton, The Bitter Lesson Hand-designed interfaces are what we have discovered. Agents can read their own trajectories, spot whatโ€™s missing, and write and test code. ๐—œ๐—ป๐˜๐—ฒ๐—ฟ๐—ณ๐—ฎ๐—ฐ๐—ฒ ๐—ฑ๐—ฒ๐˜€๐—ถ๐—ด๐—ป ๐˜€๐—ต๐—ผ๐˜‚๐—น๐—ฑ ๐—ฏ๐—ฒ ๐—ฎ๐—ป ๐—ฎ๐—ด๐—ฒ๐—ป๐˜ ๐—ฐ๐—ฎ๐—ฝ๐—ฎ๐—ฏ๐—ถ๐—น๐—ถ๐˜๐˜†. Thatโ€™s what Agent S4 does. Agent S4 makes a ๐—ฟ๐—ฒ๐—ฝ๐—ผ๐˜€๐—ถ๐˜๐—ผ๐—ฟ๐˜† ๐—ผ๐—ณ ๐—ถ๐—ป๐˜๐—ฒ๐—ฟ๐—ณ๐—ฎ๐—ฐ๐—ฒ๐˜€ ๐—ฎ๐—ป๐—ฑ ๐˜‚๐˜€๐—ฎ๐—ด๐—ฒ ๐—ธ๐—ป๐—ผ๐˜„๐—น๐—ฒ๐—ฑ๐—ด๐—ฒ the unit of evolution. A task agent attempts synthetic tasks with no success labels or verifiers. A meta-agent studies the trajectories, forms hypotheses about which actions, observations, or knowledge would help, and builds descendants through mutation or crossover. Validation tasks the meta-agent never sees decide which repositories survive. The model and harness stay fixed, and the meta-agent ๐—ป๐—ฒ๐˜ƒ๐—ฒ๐—ฟ ๐˜€๐—ฒ๐—ฒ๐˜€ ๐—ฎ ๐—ฏ๐—ฒ๐—ป๐—ฐ๐—ต๐—บ๐—ฎ๐—ฟ๐—ธ ๐˜๐—ฎ๐˜€๐—ธ. ๐—ข๐—ป๐—น๐˜† ๐˜๐—ต๐—ฒ ๐—ถ๐—ป๐˜๐—ฒ๐—ฟ๐—ณ๐—ฎ๐—ฐ๐—ฒ ๐—ฒ๐˜ƒ๐—ผ๐—น๐˜ƒ๐—ฒ๐˜€. With the evolved repository, Qwen3.8-27B goes from ๐Ÿฑ.๐Ÿฒ% to ๐Ÿฎ๐Ÿญ.๐Ÿฏ% on OSWorld V2, ๐Ÿฏ.๐Ÿดร— the same model with fixed interfaces and a new state of the art for open-source computer-use agents. It beats GPT-5.5 (13.0%, $25.50/task) and edges past Opus 4.8 (20.6%, $72.40/task) at $๐Ÿญ.๐Ÿญ๐Ÿฌ per task. On ScienceBoard, the same model goes from 41.4% to 58.6%. The analysis is the more interesting part. Evolution mostly deepens interfaces the agent already has: 4 new interfaces and 127 extensions. Give the meta-agent a scratch VM to test its code, and it builds 12 new interfaces and 253 extensions, reaching music scores, audio projects, and 3D scenes. The best repository came from crossing two earlier branches. In one case study, the seed agent hand-builds a browser client mid-task and debugs through failure after failure; the evolved agent calls page_text and reaches the same message at tool call ๐Ÿญ๐Ÿญ instead of ๐Ÿฏ๐Ÿณ. Human-like computer use was a starting point, not the destination. The interface an agent acts through is a second axis of improvement, beyond better models and harnesses, and agents can now design it themselves. ๐—˜๐˜ƒ๐—ผ๐—น๐˜ƒ๐—ถ๐—ป๐—ด ๐—ฎ๐—ด๐—ฒ๐—ป๐˜ ๐—ถ๐—ป๐˜๐—ฒ๐—ฟ๐—ณ๐—ฎ๐—ฐ๐—ฒ๐˜€ ๐—ถ๐˜€ ๐—ฎ ๐—ป๐—ฒ๐˜„ ๐—ฑ๐—ถ๐—บ๐—ฒ๐—ป๐˜€๐—ถ๐—ผ๐—ป ๐—ผ๐—ณ ๐—ฎ๐—ด๐—ฒ๐—ป๐˜ ๐˜€๐—ฒ๐—น๐—ณ-๐—ถ๐—บ๐—ฝ๐—ฟ๐—ผ๐˜ƒ๐—ฒ๐—บ๐—ฒ๐—ป๐˜.
27
40
9
138
10,474
WebAgentlab retweeted
Computer use and browser use toolsets are now built into the Python and TypeScript SDKs for Claude. The API tells you what Claude wants to click or type. Previously, you had to write your own loop and map clicks and keystrokes to commands, but the SDKs now run the loop and send actions to drivers.
135
298
61
5,314
448,418
incredibly based that Xiaomi MiMo released their RL dataset link: huggingface.co/datasets/Xiaoโ€ฆ
13
103
7
1,236
73,144
WebAgentlab retweeted
1/ Today we're announcing the stable Cua Driver release for Omarchy - a new foundation for computer use, built into the OS from the ground up. Over the last month, we worked directly with @dhh, @SpencerGBull and @vaxryy to bring a native synthetic cursor to Omarchy's Hyprland compositor, enabling true multi-cursor computer use at the OS level. Cua Driver is open source and available at github.com/trycua/cua
117
167
57
1,964
343,191
WebAgentlab retweeted
1/ You can now run OSWorld tasks on Cua Cloud Fleets. Our new guide shows how to turn the official OSWorld Ubuntu image into a Fleet desktop with Cua Driver, then run your agent and evaluate the tasks. Get started: run.cua.ai/?utm_source=x&utmโ€ฆ
1
3
1
20
6,445
WebAgentlab retweeted
There is no benchmark tells you if your agent could resolve your mechanical CAD design in accurate, editable, reliable, efficiently. We designed the benchmark CADWORLD from professional prospective that really focused on the engineers workflow. arxiv.org/abs/2609.16251
1
1
5
1,600
WebAgentlab retweeted
Check out my first music video ๐Ÿ–๏ธ Plus, accepting applications for hot (robo)secretaries Want yours?
99
45
39
295
142,932
WebAgentlab retweeted
We just raised $5.7M for @PolarBrowser, the AI browser that beats Anthropic and OpenAI on every major web agent benchmark. - 4,500,000+ actions taken for users, automating sales, recruiting, and ops - One company cancelled Clay and saves 25+ hrs/wk per person - Team is from MIT, YC, Prod, Citadel, Jane Street, Perplexity, Modal, and Apple 100 hours of work. From 30 seconds of typing. Download the world's most powerful AI browser: polarbrowser.com
1,340
1,267
194
16,334
22,505,283
New blog post: What's the Point of Computer Use Agents? There's been a lot of hype about CUAs recently. But where do CUAs shine? When should we use CUAs over API- or text-based agents? And what are the remaining research problems? jykoh.com/blog/whats-the-poiโ€ฆ
14
39
10
195
27,578
WebAgentlab retweeted
Introducing ApprenticeBench: computer use + continual learning on a real job. We show Fable 5.1 and GPT-6 Astra can now continually learn on a job and surpass human professionals. A decisive step change in AI's job readiness. No FDEs. Agents deploy themselves into the job. ๐Ÿงต
35
167
51
1,225
233,588
WebAgentlab retweeted
Watch Fable 5.1 agent onboards itself in the accounts payable department at a simulated California-based construction company. Learn the ERP system, conventions from company history, and lifelong learning from feedback. A glimpse into what the future of work may look like.
1
8
4
109
24,284
WebAgentlab retweeted
1,965
3,428
2,449
38,681
11,424,743
WebAgentlab retweeted
35 of the 88 RSI-Exam tasks are now public! ๐ŸŽ‰ Weโ€™re continuously expanding RSI-Exam and always welcome new contributors across fields. Bring a challenging task from your own research area: ๐Ÿ”— rsi-exam.ai/contribute.html
Can frontier AI agents achieve Recursive Self-Improvement by turning a weak method into one that performs better on hidden data? Introducing RSI-Exam ๐Ÿงต: 88 executable research tasks across 6 domains, now open for task contributions from every field๐Ÿง‘โ€๐Ÿ”ฌ. Its task bank spans virtual cells, TPU kernels, chip design, quantitative finance, agent harnesses, model distillation, and much more. Across these tasks, every rollout follows the same core protocol: ๐Ÿ” An agent inherits a working method and iterates using visible data. ๐Ÿ”’ Only its final artifact enters a fresh-container evaluation on the hidden set. ๐Ÿ“Š The result so far: Opus 5 leads the 88-task leaderboard with a mean hidden-set score of 0.464. ๐Ÿ“ข Want to help build the next release? Author an executable task, review one, or audit a rollout. Substantial contributions qualify for paper authorship. Special thanks to @XinyeYee for the tremendous effort in leading this project. Explore: rsi-exam.ai/ Blog: rsi-exam.ai/blog.html Contribute: rsi-exam.ai/contribute.html GitHub: github.com/aiming-lab/RSI-Exโ€ฆ
5
22
3
163
26,142