Flat versus tree, hands on

Interactive demo · illustrates Paper 1 §4, §6, and Appendix A

The same 50 skills, twice. Left: the shipped default — a flat library, every skill's cost charged before you look at any of them. Right: the same skills as a coordinate-addressed tree — you pay per node you enter, and only for the path you walk.

The task (both panes): find and load the ONE skill for: “clean a CSV export.”

Flat — the shipped default

context: 7,500 tokens

All 50 skills × ~150 tokens each were loaded the moment this pane existed. Scroll, scan, click the one you think is right — the meter was paid up front either way.

Tree — descend on demand

context: 150 tokens

Only the root index is loaded (~150 tokens). Click a node to “Read” it: that injects its layer — its children appear, the breadcrumb grows, the meter counts the node you entered. Nothing you don't enter is ever loaded.

    Flat paid for all 50 before you looked; the tree paid for the path you walked. flat: 7,500 tokens (50 skills, up front) · tree: 150 tokens (1 node entered)

    What the meters are counting

    Flat exposure costs O(#tools): every definition is in context before the first request, so the bill grows linearly with the size of the library whether or not a skill is ever used. Descent costs O(depth): one node's layer per level walked, bounded by the depth of the tree rather than the size of the library. Reaching csv-cleaner here takes four nodes (0 → 0.1 data → 0.1.1 tabular → csv-cleaner) ≈ 600 tokens, against 7,500 paid up front on the left. The argument and its calibration against the platform's own published measurement are in Paper 1, §4 and Figure 1; why a depth-one skill library is a flat forest is §6.

    Honest note: the token numbers here are illustrative of the loading model (a uniform ~150 tokens per node), not measurements. The real mechanics — coordinates standing in for identity, breadcrumbs for edges, and the platform-verified load semantics (the Read tool injects one layer; a shell cat injects nothing) — are documented in the paper's Appendix A.2.