Why is finding distributed behavior code so hard?

Before a developer or coding agent can change how a harness behaves, they have to find every code location that implements the target behavior. That sounds like ordinary code navigation, but it is uniquely hard for harnesses: production harnesses are large, tightly coupled, and behaviorally distributed — a single behavior is smeared across files, functions, execution stages, and state transitions. Meanwhile modification requests are phrased in terms of what the system should do, and repositories are organized by files, functions, and modules. There is a representational mismatch between the request (behavioral) and the artifact (structural), and closing it is left to the human.

Naming this gap matters because it relocates where the difficulty in harness evolution actually lives. Existing tools — code search, repo indexing, long-context processing — make code easier to inspect but still leave the developer to recover the behavior→implementation mapping themselves. So "harness evolution" is bottlenecked less by generating a good edit than by locating where the edit must go.

This refines the vault's picture of agentic code. Since What are the three distinct layers of agent code?, the harness layer is precisely the part that grows tangled and distributed as it accumulates behaviors — which is why it, not the model, becomes the maintenance bottleneck. The proposed remedy is to reorganize repository knowledge around runtime behavior; see Can explicit behavior maps help weaker planners compete with stronger models?.