Pragmatic Engineer ran a piece on April 24 that gave a name to something we have been watching for a while. Tokenmaxxing. Companies are starting to track AI token usage as a productivity metric. Workers are gaming it the way workers always game metrics. Parallel agents that produce more tokens than the work requires. Throwaway prompts that exist to spike the count. Real production outages from AI overuse, because the path of least resistance for an engineer being measured on tokens-burned is to write the prompt that burns the most tokens. The metric works. The work does not.
This is not a new thing. It is the oldest thing. We are watching every metric mistake of the last forty years happen again, on a one-year compression. Lines of code shipped. Bugs closed. Story points completed. Each one ran the same arc. Pick a number that loosely correlates with productivity. Promote the people who maximize it. Watch the number stop correlating with productivity, because the number is now the thing being optimized, and the work is not. Goodhart's Law, faster.
What is new is the unit of measurement. Tokens are denominated in a currency that is also a cost. Every token someone burns to look productive shows up on the company's invoice from Anthropic, OpenAI, or Google. Every metric in history was theoretically free. This one is not. The denominator is going to wake up CFOs to a problem that was always there.
Why this matters for the way you should be working
If you are reading Defrag, you are probably not in a company that explicitly tracks tokenmaxxing. You also are not exempt. The same dynamic is showing up in indie practice, where the metric is followers gained per week, repos starred, content shipped per day. The denomination is different. The structure is the same. Pick a number that is downstream of the actual work, optimize it, watch the actual work get worse. The discipline is to refuse the metric and run the loop.
Anthropic shipped task budgets on Opus 4.7 the same week the tokenmaxxing piece dropped. That is not a coincidence. Task budgets are the architectural answer to the dynamic Pragmatic Engineer described. The right amount of compute for a job is the smallest amount that produces the output, not the largest amount the budget will allow. Setting a budget per agent type and letting the agent self-stop is what tokenmaxxing isn't. We covered the budget mechanics in the Opus 4.7 task budgets piece this week.
The deeper pattern
Anthropic also shipped what some are reading as a model-tier signal in the same window. The Project Deal pricing structure, where higher tiers get access to more tokens per request, frames token consumption as a privilege class. This is the corporate version of the dynamic. Tokens are now a status good as well as a unit of work. The companies offering them have a structural interest in their consumption increasing. The companies tracking them have a structural interest in measuring it. The workers in between are stuck.
The way out is the same way out as every previous metric trap. Refuse the metric. Pick a different denominator. The Defrag loop, which we documented in the hub article, uses one specific number as the only metric that matters. One thing installed per week, taken from a brief that ran against a written scope. The token count of running the loop is irrelevant. The brief that came out of it is irrelevant. The only thing that matters is whether the install happened, and whether it changed how the next week's work felt.
What to actually do
Do not read this as the prescription. Read it as the diagnostic. The right response to tokenmaxxing as a trend is not a private vow to write fewer prompts. It is to ask the harder question, which is whether the metric you are personally optimizing for is downstream of the work or the work itself. Most metrics that feel productive are downstream. Most of them, if you keep optimizing them past a point, will hollow out the thing they were supposed to measure.
Set a task budget on your subagents. Run the loop. Install one thing this week. Skip the part where you watch the token counter go up.