Agent value can’t be measured by tokens — build verification first

The strongest signal from this talk is that agent adoption is now a verification problem, not a generation problem. Teams can generate code at machine speed, but human code review remains the bottleneck, and Anthropic's own acknowledgment that code review is unsolved underscores the gap.

The speaker argues that most organizations measure agents in tokens, which are an output, not an outcome. Borrowing from James Watt's horsepower, he makes the case that agents need a tangible ROI metric for non-early adopters; otherwise, teams oscillate between overspending on tokens and an austerity pullback.

The useful part is a task-selection matrix: low uncertainty in steps should be scripted, high uncertainty is likely out-of-distribution, and the sweet spot is tasks that are moderately uncertain but easy to validate. Those NP-style tasks justify a doer agent plus a verifier agent.

This is a thought starter, not proof. Evidence is anecdotal, and "mouse power" is explicitly a metaphor. Engineering leaders should still act on the core claim: define acceptance criteria and verification before spending on agent tooling, because unmeasured agents become unmanageable.