The Agent Parkinson's Law: Double the Budget, No Extra Work Done

The horizontal axis is the budget allocated to the agent, and the vertical axis is the tokens it actually burns. The chart splits consumption into two parts: one called USEFUL WORK, and the other called ELABORATION.

In the first half of the curve, useful work rises with the budget. After a certain point it flattens out; as the budget keeps increasing, almost all of the added amount goes into elaboration. In the upper-right corner there is an orange circle containing the text: the more budget, the more tokens are spent; more tokens do not mean more work is done.

This chart comes from tokenbender, captioned 'the current state of agent research/software engineering.' The chart isn't complicated; what's brutal is that it draws out something everyone knows but rarely puts on the table.

The comments then pile on. One user asked which part of current agent research is the biggest mind game. tokenbender replied: everyone is talking about scaling agents into a society. I hear it, and the secret is error-decoupled agent scaling; unfortunately, that isn't in your API.

To translate: a capability that hasn't made it into the API yet is being spread as a roadmap. Multi-agent, agent society, collective collaboration—these words first become true in slides, and then engineering is left to catch up. Whether it can catch up is another matter; the budget gets burned first.

Another user, aespa(x), replied more directly: discussion always revolves around doing more work, and rarely around doing better, higher-quality work.

This hits the soft spot of how things are billed. Tokens can be counted; quality cannot. So agent progress easily gets written as number of calls, context length, number of parallel agents. All these metrics go up, and it looks like progress. As for whether the output has actually gotten better, whoever accepts the work has to find separate evidence.

The original Parkinson's Law is that work expands to fill all available time. The agent version is similar: consumption expands to fill all available budget. Long-chain planning, mutual review, retrying failures, reading the same chunk of context over and over—all of these are doing work, and all of them are spending tokens. The difference is that some of it adds nothing to the final result.

So the horizontal axis of that chart can actually be replaced with any resource: budget, concurrency, context window. The more you give, the more easily the system learns to use it all up.

Next time you see an architecture diagram for an agent society, you can first ask: what verifiable thing did that extra token buy? Budget is not a progress bar.