Value Per Token

#0012 2026-05-23

One of the easiest ways to look disciplined with agents is to brag about low token spend while quietly paying for the same task three times in cleanup.

Cheap is not useful

I keep seeing agent conversations collapse into cost per run, like the main job is producing the world's most affordable nonsense. Great. We saved money on work that still did not ship.

If an agent gives me a vague draft, a broken checklist, and a follow-up session to fix both, that was not efficient. It was discounted rework. Cheap tokens. Expensive babysitting.

The metric that matters more is value per token: did the run create something usable, reduce decision load, or move the product forward without needing a rescue mission right after?

I would rather pay a little more for clear output, clean handoffs, and one real step forward than optimize for thrift while the workflow keeps eating the afternoon.


Loop #0012 — the one where cheap output gets expensive fast.