ThoughtsJul 26, 20265 min read
The Orientation Tax
What a coding agent's first day on the job actually costs, and why the environment is the only lever on it. On the environment-performance study.
A companion to the research note How the Environment Affects Agent Performance and Token Cost. Every figure below is that study’s; this piece argues about them rather than adding to them.
A new engineer's first week is mostly not engineering. It is reading: the codebase, the conventions, the unwritten rules about what done means here. Nobody calls that waste. We call it orientation, we pay it once, and we expect it to come back. An agent pays it too, on every task, and we bill it as tokens.
The pilot behind this piece held everything still except the room. One coding agent, one software task, one codebase, one prompt, one model, and six escalating levels of project context, from an empty workspace up to a full project scaffold. The only variable was how much the environment told the agent about where it had landed.
Two things came out of it. Every environment finished the task, so richer context did not buy success; the agent could already do the work. And the scaffold was close to free to carry, because caching served 88 to 96 percent of the input. Whatever a project brief costs to write, it is not costing much to read.
The part that did not survive
The pilot also found the richest environments spending up to 36 percent fewer tokens, and that is the number this essay is not going to lean on. A later 84-run matrix across two agents and two tasks did not reproduce it. More context trended costlier there, not cheaper.
This is the honest shape of the result, and it is worth sitting with rather than skipping past. The framing survived. The headline did not. An orientation tax is still the right way to read what a token bill is partly measuring, but the claim that a better environment reliably lowers that bill was a one-task pilot talking, and the fuller study took it back.
Why the framing still earns its place
Because it points at the right lever. If some fraction of every task's cost is the agent working out what the work is, then that fraction is a property of the room, not of the model. You cannot prompt it away and you do not need a better agent to move it. It is the one part of the bill that belongs to you.
What the program has been doing since is finding out whether agents actually collect on that. The answer turned out to be more complicated than the pilot suggested, which is the subject of everything that came after it.
The study, its six context levels, the per-cell data and the caching figures are in the research note.