The 7-agent experiment confirmed it directly. Agents that wrote structured DECISIONS.md and PROGRESS.md files outperformed agents that dumped raw logs. Same models. Same prompts. Same starting budget. The variable was the thinking architecture between sessions.