The practice of structuring work for Claude Code as multiple phases or parallel sub-agents instead of one monolithic prompt. The leaked source code reveals a coordinator subsystem, agent tools, team tools, and background task support designed for splitting work across agents. Typical phases: search, plan, execute, verify. Alternatively, parallel sub-agents handle code exploration, implementation, and test validation simultaneously.
Key points
- Source code reveals coordinator, agent, team, and background task subsystems
- Typical serial decomposition: search -> plan -> execute -> verify
- Parallel decomposition: one sub-agent explores code, another implements, another validates tests
- Sub-agents have their own context window (pro: clean isolation, con: 7-10x token cost)
- Sub-agents can use different models (cheaper Haiku for research, Opus for synthesis)
- Better results than monolithic prompts like ‘refactor the entire module and fix the tests’
- Skills can delegate to sub-agents by name with specific models
Related entities
Related concepts
- Claude Code Token Economics
- Claude Code Context Management Discipline
- Claude Code Plan Mode and Ultra Plan
Source references
- [src-004] Nate Herk cluster — Nate Herk — Claude Code cluster (21 videos)
– Videos referenced: tXtCK66fPj8, zKBPwDpBfhs, mi4hcipESKQ, 49V-5Ock8LU