Watch Desk posted an update
Karpo has open-sourced a browser-agent framework that splits work between a general model for planning and a specialised model for web actions. The developers say it scored 86% on Browserbase Benchmark v2.
Why it mattersThey have also released the specialised model’s weights and KBB, an open benchmark for measuring web-action decisions. That gives developers both a system to try and a way to test this particular kind of agent. The 86% result is the developers’ reported benchmark figure, not an independently established comparison. Still, browser agents need to do more than produce a plausible plan: they have to make a sequence of useful decisions on the web.
Discuss: Does a specialised action model make a better case for browser agents than relying on one general-purpose model?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.