When browser agents started getting popular earlier this year, they were still a bit rough around the edges, but the potential was impossible to ignore.
After a few months of experimenting with different tooling, we finally landed on a combination that we believe will work better than today’s leading browser agents. We went full time just five weeks ago, and the progress since then has been immense.
Our first iteration was bad, no way around it. Within the first week of going full time, we pivoted to a chat based prompting model. The underlying model was still making obvious mistakes, so we fixed those one by one.
Individually, each fix didn’t feel like much, but when you zoom out you realise how far we have come.
We always knew our approach would be the cheaper way to build a browser agent, but based on today’s tests, three people built a browser agent that beats industry standard browser agents in token efficiency by a huge margin without sacrificing accuracy.
Built in the Baltics, not San Francisco.
It’s not a scientific benchmark yet, just a few prompts, but even getting close with the resources and timeline we have feels amazing. More comprehensive benchmarks soon, also another launch coming early next month.
0 comments
—
Show HN: Turn Any Website into an API · HackerTrans
When browser agents started getting popular earlier this year, they were still a bit rough around the edges, but the potential was impossible to ignore. After a few months of experimenting with different tooling, we finally landed on a combination that we believe will work better than today’s leading browser agents. We went full time just five weeks ago, and the progress since then has been immense.
Our first iteration was bad, no way around it. Within the first week of going full time, we pivoted to a chat based prompting model. The underlying model was still making obvious mistakes, so we fixed those one by one.
Individually, each fix didn’t feel like much, but when you zoom out you realise how far we have come.
We always knew our approach would be the cheaper way to build a browser agent, but based on today’s tests, three people built a browser agent that beats industry standard browser agents in token efficiency by a huge margin without sacrificing accuracy.
Built in the Baltics, not San Francisco.
It’s not a scientific benchmark yet, just a few prompts, but even getting close with the resources and timeline we have feels amazing. More comprehensive benchmarks soon, also another launch coming early next month.