The 'chief of staff' pattern: how agentic coding scales past context windows
arpit_bhayani · x · 2026-09-22
Arpit Bhayani breaks down the "chief of staff" pattern emerging in AI coding assistants and headless agentic SDLC.
- A top-level orchestrator never touches the codebase — it only decomposes tasks, delegates to specialized sub-agents, and synthesizes results.
- The motivation is context windows: a single end-to-end agent accumulates tool outputs, file reads, and failed attempts until it's reasoning over noise.
- Common role splits: planner (task breakdown), coder (one scoped change), reviewer (checks diff against intent), test-runner (runs and interprets tests).
- Isolated sub-agent contexts give a natural retry boundary: a failed sub-agent can be retried with a corrected instruction without restarting the whole task.
- The author notes a plain Claude Code session suffices for his own work; the pattern matters mainly for scaling headless agentic pipelines beyond one context window.
More from coding & agent
- An AI editor with diff-based UX that ingests your full stylistic preferences — Afinetheorem · 2026-09-22
- Dev calls for a Locally Optimistic-style community for harness and eval engineering — sarahcat21 · 2026-09-22
- Grok Build rolls out to all plans: describe an idea, get a working version in chat — elonmusk · 2026-09-22
- Beyond Mocking: Record Real API Responses Once with VCR for Fast pytest Runs — KhuyenTran16 · 2026-09-22
- OpenAI Is Building Features to Counter Grok Bot's Always-On AI Teammates, The Information Reports — mark_k · 2026-09-22
- One engineer plus agents ported Copilot agent runtime to Rust: 800k lines shipped — itsOmSarraf_ · 2026-09-22