Tokenmaxxing Versus Token Budgeting

Right now, the CEO’s are saying, “Do more with AI! Go faster! More agents! Use as many tokens as you can!”

The investors are on board. “How much can we get done with less people? Can we build a $1B company with 3 people? I’ll keep investing in your token costs. Max it up!”

But the CFO’s are starting to wince. “We spent $1.5M on tokens last month. You — just you — spent $37k on tokens this weekend alone. What the heck were you doing? We can’t keep this up.”

I’m now convinced that the future of knowledge worker compensation negotiation contains a token budget. I don’t know when that happens, but I’m convinced it will.

In preparation, I’m starting to think about the following:

• What should we be spending tokens on?
• Which model for what task?
• How do we budget and limit our autonomous agents’ token spend?
• When should our autonomous agents use a deterministic toolset (ie, just code) vs an LLM?
• How do we track our autonomous agent token spend?

There’s a future here in how, on what, and when we spend tokens.

But for now, max it up!