But then I would expect like even just running the agent for like an hour it will accumulate a really really large number of workflows right even before you say like we go to millions of tasks we already got a lot of data then how could we effectively incorporate them into maybe the context window the model has

yeah i think that's a really interesting question but that's also what i was trying to do but i feel like for now maybe even, that the benchmarks where the data sets of agents that we have doesn't support that scale yet.

So I think it would be interesting to actually find other sources or find more richer sources to collect a bunch of agent trajectories.

I think there are works that try to let the agent explore the environment just by synthesizing tasks themselves and to accumulate that.

I think that's one way to do it.

And for the second part, if we have that huge repository of tasks or workflows, how do you let agents to integrate those? I guess from our experiments, one is at least for a particular user on a particular website or environment, there are a number of tasks you do, and it's just being essentially with different contacts.

So actually the procedure are shared most of the time.

But how would you think like two years later, what type of question you would be working on? That's a great question.