Tokens for yesterday

Here's the embarrassing version of agentic work. You ask to move a field. The agent reads a ticket written up from a meeting, a spec that summarizes the ticket, architecture notes that contradict the code, and then finally the code itself. A review agent reads the whole pile again, and a status agent turns the result back into prose for people who will skim it. There's intelligence at every step and, as far as I can tell, nobody asking why the steps are there.

The run below is illustrative, but I don't think it exaggerates much:

request                         12 wordsticket + spec + architecture    18,000 tokensrepository context              31,000 tokensreview + repair                 14,000 tokensstatus report                    1,200 tokensvisible result                  one field moved

The waste isn't that the model talked. It's that we make it reenact a bureaucracy built to move context between humans who couldn't share it quickly — the ticket, the spec, the status report all made sense when the readers were people sitting in different meetings.

The artifact laundering machine#

Anthropic's analysis of more than four million Claude conversations found that software development and writing together accounted for nearly half of usage. Both useful domains — and both prolific producers of symbolic residue: tickets, wrappers, status reports, migration notes, summaries, explanations of explanations.

The residue becomes tomorrow's prompt. In a randomized trial, METR found that experienced developers took 19 percent longer with early-2025 AI tools on mature projects they already knew, while believing the whole time that the tools were making them faster. On a project like that, most of the work is archaeology — rereading old decisions so a new one can fit through them — and a fast generator mostly adds to the pile it's supposed to be digging through.

A bigger context window won't buy your way out either. The Lost in the Middle study found that models can struggle to use relevant information depending on where it sits in a long prompt, so shoveling the entire archive in mostly gets you a bigger warehouse to lose the box in.

Delete the task#

Before asking which model should automate a workflow, ask which parts of the workflow only survive because people needed handoffs. An agent that maintains shared state doesn't need status reports written for another agent to summarize. A user who can interrogate the source of truth doesn't need it cloned into a dashboard and three documents. And if the acceptance criteria can be executable, you can stop translating the same intent through a ticket, a specification, a patch description, and a review essay.

Some problems stay stubbornly real no matter how much process you delete — safety, law, goals that genuinely conflict, inputs you can't trust, the physical world. Stop spending the expensive tokens on proof that a model can navigate an org chart and point them at those instead.

Every agent leaves a receipt, the receipts feel like productivity, and the next agent pays for them as context — which is the quiet way AI defeats itself, producing artifacts faster than it removes the coordination that required them. The cheapest token is the one attached to a task you deleted.