Tokens Don't Make the Driver
2026-10-15 · AI-translated, author-reviewed
I don't think counting the tokens spent is a valid way to judge how well someone uses agents. In work teams, those numbers can become a sort of ranking, where whoever consumes the most sits at the top of the pyramid and wins. But alas, that's not how it works. And I'm writing this article to offer a different way of looking at it.
The race
If we take a car race as an example, the driver who burns through tyres or fuel the fastest isn't the one who gets a place on the podium, for both physical and logical reasons.
As a driver gains experience on the track, they learn about tyre grip, traction on the asphalt, and the limits of the car. Then comes throttle control too: you can't always go flat out, unless you want to end up smashing into a track barrier.
Something similar happens with agents. You can keep pressing the accelerator, adding requests, running tools, and consuming tokens. That alone doesn't tell you how much closer you're getting to the result.
Fuel and grip
So we could say that tokens are the fuel and context is the tyres.
Context gives you grip on the problem: knowing where to work, which constraints to respect, and what it means to be finished. Without that grip, you can have all the fuel you want and spend your time correcting your line. And adding random context doesn't necessarily solve the problem: you need the relevant context.
Consuming less doesn't automatically make you a better driver either. If you save fuel but never reach the finish line, there's little to celebrate.
Before the podium
Before looking at how many tokens were spent, I'd look at what was achieved, whether it works, and how much work it took to check and correct it. Consumption is still useful for understanding the cost. On its own, it tells us very little about how good we've become.
From here, I leave all further speculation to you, the reader. But before stepping onto the podium, let's at least make sure we've finished the race.