A single conceptual edit forces the agent to identify, load, and modify many small sites — each cheap individually but the search and completeness-check cost dominates. The agent's context-window load rises across N files just to verify the same logical change everywhere, and the retrieval cost of finding the next site competes with reasoning about the change itself.
All code that varies together sits in one place; the agent loads one module to make any change along this axis and verifies completeness in one read. The agent's context window holds only the relevant slice; the reasoning step count drops because the agent doesn't reconstruct the scattered set, and completeness-check cost collapses to a single file.
Smellier version
// Five files each have:log(`event=${event}, user=${user}`);
Fresher version
// One file:function logEvent({ event, user }) {// one place to evolve}
Every change carries risk of missing a site the agent didn't grep for; reviewers (human or agent) can't verify completeness without re-running the same search, and the agent's verification-surface cost multiplies by the count of touched files. The agent generates an edit that targeted every site it found while missing the one site the search didn't surface.
Consolidation creates a new boundary the agent must respect; previously-independent sites now route through one module that becomes a contention point for unrelated edits. Context-window load on the consolidated module rises because every change along this axis loads its full surface; the agent's retrieval cost on cross-cutting edits also rises along the new seam.
A change to the consolidated behavior lands at one file; the agent's token cost per edit drops from N files loaded to one, and the chance of missing a site goes to zero by construction. Verification surface collapses because the file is the surface; the agent reads one module and ships the change without grep-and-iterate.
Pulling every superficially-related edit into one module creates a god-module the agent now loads as a tangle of unrelated concerns — the smell migrated, not vanished. The agent's reasoning-step cost rises as it disentangles unrelated axes per edit, and context-window load on the god-module exceeds what the agent can hold for any single concern.