- Published on
Orchestrating Claude Code Agents: The Chief of Staff Pattern
- Authors
- Name
- DeepHorus
A new orchestration pattern for AI coding agents, called the Chief of Staff pattern, separates orchestration from execution. One long-lived coordinator session assigns work, verifies claims, and maintains shared state, while short-lived sessions implement. It borrows from Git's integration-manager workflow, where a maintainer tests changes before they land.
The problem: a single AI coding session works well for about an hour, then degrades. Context is finite and lossy, self-reports drift, and lessons do not compound unless written down. Adding more agents multiplies the problem, creating unreliable reporters with no one reconciling them.
The pattern is also called orchestrator-worker, coordinator-implementor-verifier, and maker-checker. The coordinator writes briefs, verifies claims, and reads diffs, but never implements. State lives in a durable external board or task system with an API, surviving compaction, session death, and handoffs.
Every agent report is evidence, not instruction. Commands are re-run; exit codes are authoritative, summaries are intent. Messages go to durable channels like committed files or board cards, because in-memory messages can be delayed or expire.
This discipline separates prototypes from shipped products. Generation was never the bottleneck; checking is. The most expensive errors come from checks that report success for work they did not do. Timeboxing should surface progress, never cut work.
The name "Chief of Staff" is a metaphor, not an established term. Claude Code's subagent documentation frames the same shape as team lead and teammates. The fix is organizational, not technical: someone's job is to know what is true, not to do the work.
As AI agents take on longer tasks, durable external state and verification chains will likely become standard. The pattern makes self-reports reliable and helps lessons compound across sessions. It targets the real cause of long-horizon AI coding failure: unreliable context and unchecked claims.
