Fix: Bump max_iterations when resuming due to throttling

[tofarr]

This is intended to address #3323

There was a bug where when you reached the max_iterations, the UI would keep asking you to resume after every single request. This resets the iteration count. (100 by default - so you get 100 more requests before the session pauses again)

Testing

I didn't want to make 100 requests each time, so I temporarily updated the max_iterations in my local config.toml from 100 to 2. ;)

Sample Output

I asked the agent to write a couple of poems - it now resets the iteration count when resuming, so that you can hit max_iterations (In my case 3) before the agent stops again:

[enyst]

I think this place in the code is also hit when max_budget_per_task is exceeded, and the user wants to continue. In that case, to reset iterations doesn't seem right.

[tofarr]

Would something like the following make more sense:

            # Reset the iterations (So user will not be immediately asked again)
            if self.max_budget_per_task is None or self.state.metrics.accumulated_cost < self.max_budget_per_task:
                self.state.iteration = 0

Or do you think it would make sense to expand the max_iterations instead (Though this could result in balooning of task size)? I suppose we could even have an explicit reset iterations action.

[enyst]

It should work. It could use a bit of testing, I guess I'd suggest also a test with a delegate agent getting this iterations issue, just in case, but it looks good to me. Cc: @li-boxuan

[li-boxuan]

Expanding max_iterations seems to make more sense than resetting self.state.iteration. Resetting self.state.iteration will make backend output look weird.

Another issue, as pointed out by @enyst , is we also need to take care of budget control. Maybe we should also expand the budget (rather than simply resetting it).

After all these, you could delete all TrafficControlState related code.

[li-boxuan]

This piece of code alone doesn't work in a multi-agent context. This essentially expands the current agent/controller's limit, but when it returns to its parent, the old limit would be hit immediately. It would still work but a little bit confusing/surprising to users. The backend logs would look weird, too.

I would suggest take a look at code in if delegate_done section. You probably need to do something similar to "propagate back" the max_iterations and max_budget_per_task.

[tobitege]

@tofarr I hope you don't mind that I tried fixing the lint error and main merge, but then mypy also tried to reformat the file again and caused that as an error... 🤯 That thing annoys me.

[tobitege]

Instead of multiplying the max_iterations, now just add the initial max_iterations to it (the param that originally got passed into AgentController).
Tested locally and works.

[tobitege]

At least from my point of view, this looks good 😁
Thanks, Tim!

[enyst]

@li-boxuan has pointed out here a weird consequence with delegates: if the user asks for, lets say, searching for info on the web, an agent will delegate to another, and if this 'child' agent, the browsing agent, hits max_iterations, then:

• the user can choose to continue, at, lets say, 100/100
• the browsing agent does more stuff
• it ends, and control returns to the 'parent' agent
• which will stop immediately with error, because the parent state wasn't updated, so the user sees it's at 178/100 or 102/100, whichever it got to.

If that's the case indeed, it will look odd.

I would suggest to replicate it and see. The user can say continue again, I assume, I just suggest to give it a try and see how it feels like, and how you think about it.