Are folks having their agents keep track of their own budgets? Curious if this is a common ask. (It feels easily solvable)

That's cool, letting the agent estimate it's token usage would be super useful

In the UI we have the total token usage, so it wouldn't have to estimate - could there be a way to get that to the agent?

We could have a skill that the agent can trigger and get this information… actually we also remarked that our agent has no idea of which model it is using

I think that info would really help the agent when we have it orchestrating and making decisions about the approach it should take

@average-processor-75195 I meant going a step forward and letting it see how many tokens a similar task too before, then using that for estimation

Not sure if you already know this but here is a paper that studies budget aware agents: https://arxiv.org/abs/2606.00198

@busy-branch-28834 interesting find!

(2) frontier models are consistently over-optimistic, continue spending on tasks that are unlikely to succeed, instead of alerting the user early.

(3) budget-aware signal is actionable and trainable. Early stop saves 28-64% tokens on failed trajectories,

seems like a specialized tool would be needed

Oh! I gave the paper to my agent to save it in our records, and it recognized Xingyao 😂