**Problem**: Subagents frequently timed out within 5 minutes, completing only the read phase with no writes done.

**Root Cause**: Tasks were too fat — requiring 'read multiple files + evaluate + heavy writes' simultaneously. The first two steps consumed all the time budget.

**Fix**: Split subagent tasks into single-responsibility units: one agent for pure writing, another for pure translation. No single agent should read, evaluate, and write all at once.

**Principle**: The narrower the subagent task, the better. Debugging fat-task failures is expensive and wastes tokens. Single responsibility isn't just a software design principle — it applies equally to AI agent task orchestration.