Companies are pulling back on AI spend for the first time.
We ran a controlled experiment to answer one question:
Does structured documentation actually make coding agents better?
We worked with a company running Claude Code on million-line monorepos daily, with a full team dedicated to one thing: making their agents more efficient.
We tested their environment against two conditions:
1. No docs, just the codebase
2. Docs the codebase
When you provide agents with structured information, you get
→ 64% more precise answers
→ 39% better discoverability
→ ~50% fewer tokens per task
→ 1.5x faster completion
Better models don't address the issues on cost or efficiency. Working from raw code, they still relearn the whole codebase every session, burning context just to orient themselves.