What if the biggest lever on agent cost is not the model, but what you let it read?

The image below is a code map rendered natively in the plugin I work in. 1974 files, 17854 symbols, grouped by folder. Every circle is a boundary, every dot a tracked symbol. It is the whole estate at a glance.

Here is why it earns its place in the workflow.

Point an agent at a repo this size and it rebuilds its mental model from scratch every turn. It reads whole files. It greps, reads, greps again. Most of those tokens buy nothing, because the agent is re-learning structure it half knew a moment ago. You pay that tax on every turn.

A map like this changes the starting position. The agent gets the compressed structure first. Boundaries, symbols, how the parts relate, what each region is for. It reasons over that, then pulls full source only for the handful of places that matter for the task in hand.

The visible win is token usage, and in a codebase this size the saving compounds fast.

The quieter win surprised me. Even when the whole thing would fit in context, the agent reasons more clearly over a curated map than over a wall of undifferentiated source. Deciding what it sees is not just cheaper. It sharpens the thinking.

Two things I had to get honest about.

A map is only as good as its staleness discipline. One that drifts from the source is worse than none, because the agent reasons confidently over a structure that no longer exists. A timestamp that reads scanned at 11:32, no boundary drift, is not decoration. It is the map telling me it still matches reality, and it is the difference between a tool I trust and a picture I hope is true.

And compression has a resolution floor. Some tasks need the body, not the outline. Control flow, error handling, the odd conditional someone added at 2am. The craft is in the agent knowing when the map runs out and reaching for source, rather than filling the gap with a guess.

That is the part that stayed with me. Working this way turned “what does the agent see” from a default I never questioned into a decision I make on purpose.