Agents need to know in real-time what the current utilization is. I went through 10 iterations of this tool - and likely going to do a number of more iterations. The net net is that I needed something to show me if I am on pace to use the tokens alloted to my subscription.
Using up the session (maxxing out my session) is not the right way because you run out of tokens by Wednesday at 6pm roughly. So how do you tell your agents that are running in the middle of the night to "not use too many tokens"?
So the solution I found was a 5hr ROLLING window. Tokens are allocated across the 7 day period as if they were in chunks. Say 100k tokens an hour for easy math.
500k tokens per session, 10 sessions left in the period. 5M tokens remain.
Now - the software updates each second. Based on your usage in the window, you have this much left to spend. It snapshots your availability at any time. Using less in the last 10 mins, gives you more to use in the next 10 mins.
Now I give agents access to /maxx --session, so if n number of agents are running they can hit the session and they dont have to plan centrally for token management.
If they have been allocated by central command, and there is token availability right now, it can proceed.
Still a work in progress, but Sunday was the last straw. Sub agents blew 70M tokens with nothing to show for it, and I had a full day of client work (Monday) that didnt get done because I hit the week wall a day early.
There are some goodies in there like -
By model weighting Token calculation on a weekly basis (its a guess of course because claude doesnt tell me) Weekly pacing bar Color coded session bar if you are running hot, etc.
Ping me with questions, happy to help. MIT - I know there are issues with the software, but its been useful already. Havent hit session limits, weekly limits stable. Some agents hitting /maxx and slowing down. This could work :)
x @reif_tauati