From MCode Engineering Notes 01. Skill version: skills/subagent-split-check.
- Will the intermediate output of this subtask be needed later? Only split if you just need the conclusion.
- The reason for splitting is context, not "this belongs to another team" or "this is a different kind of question".
- There is eval data showing the single loop really gets worse, not just a feeling.
- Are there more tool definitions than the model handles well? Merge by user intent first and defer rarely used tools behind tool search.
- Procedures and rules are skills, actions and live data are tools, reference material is loaded on demand.
- The plan, decisions and review conclusions are written into a design or progress doc, not only in the conversation.
- An end-to-end eval was run before and after, flipped tasks were inspected, and each task was run more than once.