Output Over Token Spend

Tokens are an input

By mid-2026, token budgets had become a management obsession inside many engineering organizations. Scott Wu's response separates the underlying instinct from how it was implemented: "I think it is directionally correct, but there are definitely some places where people have gotten carried away. People talk about, oh, like yeah, like we rank our engineers by how many tokens they're spending. Well, let's try and rank people by how much output they're actually producing, or how much good work is actually getting done."¹

His arithmetic is deliberately blunt: compute is expensive, but if engineers ship three times more using it, the spend is very clearly worth it, so the spend question is settled by the output question and the thing worth measuring is the output. What he substitutes for a per-engineer token count is a project-level metric instead, which tickets closed, which initiatives actually moved, and above all the counterfactual case: a project scoped at eighteen months and priced at fifteen million dollars for an outsourced contractor, completed internally by the customer's own team for one million dollars in three months. That is a procurement comparison, not a productivity statistic, and it is the number Wu wants held in mind when the return on this kind of spend is being judged.

The live disagreement

This sets up one of the sharpest disagreements between two founders describing adjacent visions of agentic work. Eric Glyman's position is that tokens are a third mega-category of business spend alongside people and vendors, distinguished from other software costs by carrying a real marginal cost per job done, and that this category therefore needs a management layer: see it, understand it, control it, and route the easy work to cheaper models.² Wu agrees entirely with the routing mechanism and disagrees just as clearly with the measurement culture that grew up around it: a token count tells an organization what was consumed, not what was produced, and an organization that ranks engineers by consumption has quietly installed an incentive to consume more, not less.

The two positions are reconcilable at the level of the line item deserving visibility, and genuinely opposed at the level of what the number is actually for. The likely resolution, stated by neither man directly, is that token spend functions well as a cost-control instrument and poorly as a performance instrument, and the failure Wu is describing is an organization that confused the two simply because the cost instrument happened to be the one that shipped with a dashboard. It is also worth noting who is arguing which side: the company selling visibility into token spend has an interest in the number mattering, while the company selling the finished outcome and absorbing tokens into its own margin has an interest in the number disappearing from view. Neither party is disinterested, and the disagreement is more useful for it.

What remains unresolved

Wu's own replacement metric is asserted rather than fully worked out: measuring individual engineering output is a decades-old unsolved problem, and his own supporting examples are all project-level rather than per-engineer, which is precisely the level of granularity a token dashboard already delivers. Output measurement also carries its own version of the same gaming risk, since tickets closed is not obviously harder to game than tokens burned. And Glyman's underlying case for the token line item, that a finance team needs to allocate the cost between operating expense and research and development, by team, with a clear return, is not addressed by shifting the conversation to output at all; the allocation problem the token framing exists to solve stays unsolved either way.