what would you actually train a browser-agent model to be good at?

i’ve been thinking about browser agents less as “llms that can click things” and more as a pretty weird sequential decision-making problem. getting the first few actions right usually isn’t that impressive anymore. the harder stuff seems to be: - noticing that an action silently failed - recovering without repeating the same thing 8 times - remembering what matters from 20 steps ago - handling unexpected page states - knowing when to backtrack vs keep going - deciding when the task is actually complete so if you had a giant dataset of browser-agent trajectories, what behavior would you optimize for? raw task success seems obvious, but it feels like that misses a lot. two agents can both fail a task, but one realizes it’s stuck after 2 steps while another burns 50 actions first. we’ve been thinking about this while working on mako at tinyfish, a model specifically for web agents, and i’m curious what people here would actually want reflected in the training objective/evals. what’s the browser-agent behavior you’d most want a model to learn that current models consistently suck at?