What’s still hard about controlling agent spend?

General
1

For people running agents that call paid tools.

What’s still painful about spend control?

Not the ideal setup the part that still breaks or stays manual.

Examples:

  • one agent looping and burning budget
  • no hard per-agent limit
  • hard to see what each agent spent
  • no clean way to pause one agent fast

What’s the messy part in your current setup?

2

The messy part for me was always attribution inside the agent tree. I’ve built multi-agent orchestration setups for clients, and the same thing kept happening at the end of the month: the bill showed a total, but nobody could say which agent in the chain burnt it. We tracked it with Excel sheets and shell scripts, and even then we only had an after-the-fact audit trail, never real control.

In CrewAI specifically, max_iter and max_rpm bound loops and call rate but not dollars, so one agent dragging a big accumulated context can still get expensive while staying inside its limits. And a per-agent cap only really works if the check happens before the call goes out, not in a callback after the response comes back.

Curious what prompted the question, are you hitting this in your own crews or building something for it?