Eric Zeng @ericzeng_builds - Twitter Profile

Pinned Tweet

9 days ago

i'm incredibly pleased to finally share what we’ve been building with OpenSwarm Browser Use. the goal was never just to make an agent that can click around a website. we want to build toward the actual Jarvis future: software that can understand what you’re trying to accomplish, navigate the web, use the tools you already use, coordinate work across agents, and eventually automate huge parts of everyday knowledge work from end to end. browser use is one of the most important pieces of that. today, our research system is reaching: → 91.0% on BU Bench V1 → 94.1% across 3 MiniWoB++ seeds — 353/375 episodes solved → strong results across Odysseys → real-world tasks across GitHub, Google, LinkedIn, Gmail, Reddit, YouTube, Amazon and more → internal tests built around actual knowledge-work workflows instead of only toy browser environments and one of the results I’m most excited about: using the same recorded Opus 5 model alias on MiniWoB++, OpenSwarm reached 92.0% success versus 67.2% for browser-use. at the same time, successful runs took roughly 3x faster while using 30.2% fewer recorded tokens. across three separate seeds, that performance actually held up: 92.0%, 95.2%, and 95.2%. what makes me even more excited is where those gains are coming from. a browser agent is a whole system. it has to understand the page, find the right controls, decide what to do, interact with weird interfaces, verify that the action actually worked, recover when it didn’t, and do all of that without wasting a model call on every tiny click. we’ve spent a lot of time on the invisible parts of that loop: compact representations of webpages, action-first model responses, grouping related actions together, dedicated handling for things like forms/date pickers/dragging, and checking page state after actions. that shows up across the task distribution too. In our same-model frontier benchmark comparison, OpenSwarm led browser-use in 7 task categories and tied in 2, with some of the biggest gaps showing up in email, spatial tasks, dragging, reading and forms. and this is still early. the benchmark isn't the product and the browser isn't the end goal. the end goal is an open system where you can tell your computer what you want done and it can actually go do the work: research something, navigate tools, send emails, fill forms, collect information, operate software, coordinate other agents, produce the final output, and bring you in only when your judgment is actually needed. basically, the computer starts becoming less like a collection of apps you manually operate and more like a team you direct. that’s the future I want OpenSwarm to help build. and importantly, we want that future to be open.

ericzeng_builds's tweet photo. i'm incredibly pleased to finally share what we’ve been building with OpenSwarm Browser Use.

the goal was never just to make an agent that can click around a website.

we want to build toward the actual Jarvis future: software that can understand what you’re trying to accomplish, navigate the web, use the tools you already use, coordinate work across agents, and eventually automate huge parts of everyday knowledge work from end to end.

browser use is one of the most important pieces of that.

today, our research system is reaching:

→ 91.0% on BU Bench V1

→ 94.1% across 3 MiniWoB++ seeds — 353/375 episodes solved

→ strong results across Odysseys

→ real-world tasks across GitHub, Google, LinkedIn, Gmail, Reddit, YouTube, Amazon and more

→ internal tests built around actual knowledge-work workflows instead of only toy browser environments

and one of the results I’m most excited about:
using the same recorded Opus 5 model alias on MiniWoB++, OpenSwarm reached 92.0% success versus 67.2% for browser-use.

at the same time, successful runs took roughly 3x faster while using 30.2% fewer recorded tokens.

across three separate seeds, that performance actually held up: 92.0%, 95.2%, and 95.2%.

what makes me even more excited is where those gains are coming from.

a browser agent is a whole system.

it has to understand the page, find the right controls, decide what to do, interact with weird interfaces, verify that the action actually worked, recover when it didn’t, and do all of that without wasting a model call on every tiny click.

we’ve spent a lot of time on the invisible parts of that loop: compact representations of webpages, action-first model responses, grouping related actions together, dedicated handling for things like forms/date pickers/dragging, and checking page state after actions.

that shows up across the task distribution too. In our same-model frontier benchmark comparison, OpenSwarm led browser-use in 7 task categories and tied in 2, with some of the biggest gaps showing up in email, spatial tasks, dragging, reading and forms.

and this is still early.

the benchmark isn't the product and the browser isn't the end goal.

the end goal is an open system where you can tell your computer what you want done and it can actually go do the work: research something, navigate tools, send emails, fill forms, collect information, operate software, coordinate other agents, produce the final output, and bring you in only when your judgment is actually needed.

basically, the computer starts becoming less like a collection of apps you manually operate and more like a team you direct.

that’s the future I want OpenSwarm to help build.

and importantly, we want that future to be open.

18

43

8

7

19K

Eric Zeng

@ericzeng_builds

Last Seen Users on Sotwe

Trends for you

Most Popular Users