Can explicit behavior maps help weaker planners compete with stronger models?

If Why is finding distributed behavior code so hard?, the fix is to build the missing map explicitly rather than make an agent recover it every time. Harness Handbook is a behavior-centric representation: it links behavioral descriptions to the distributed source that implements them, constructs that mapping directly from source, and supports editing through behavior-guided progressive disclosure plus automatic resynchronization when the code changes.

The results argue that the mapping — not raw model strength — was doing the work. With the handbook, overall win rates rise 10.0 and 18.9 points on Codex and Terminus-2 while planner token use falls 12.7% and 8.6%: better plans for less compute, because the planner no longer burns tokens rediscovering where behavior lives. The sharper evidence is substitutive: with the handbook a weaker planner matches the implementation-site localization of substantially stronger models, improving every file- and symbol-level Recall/Precision/F1 comparison against two independent reference plans, and the gains persist across request types and difficulty levels.

This is a concrete instance of a recurring pattern in the harness literature — capability relocating out of the model and into the scaffolding around it. Like Can externalized bookkeeping let smaller search agents beat larger ones?, the move is to externalize a structure (here, the behavior→code map) so the model spends its budget on judgment rather than bookkeeping. It also complements What happens to code that agents create and then share?: a durable, resynchronized representation is exactly the kind of persistent artifact that compounds.