Orchestrating Claude Code Agents: The Chief of Staff Pattern

Long-running autonomous coding fails less because models can’t write code and more because their context is ephemeral and their self-reports are unreliable. The Chief of Staff pattern fixes that by separating an orchestration session (a long-lived coordinator or overwatch) from short-lived executor sessions, and by putting shared state in a durable board instead of in conversation context. The coordinator pulls tasks from the board, writes precise briefs, spawns isolated Claude Code sessions (via a session substrate like cmux), and - critically - verifies every claimed action by re-running commands and reading diffs and artifacts. The durable store (Plan Desk) holds tasks-as-contracts, atomic status flips, linked design docs and comments so lessons persist across sessions and handoffs.

The operational specifics matter: run the verifier first (a “red gate”) so a passing check proves the check actually ran; commit one task per commit so git history maps to decisions; never accept an agent’s summary as truth - treat reports as evidence, not instruction - and always verify artifacts by reading values back from the artifact itself. Watch for vacuous assertions, silent no-matches, swallowed errors, wrong references and scope mismatches; defend by making checks fail visibly, adding positive controls, using durable channels for messages, and timeboxing reporting cadence rather than work completion. This discipline converts flaky multi-agent runs into reliable, auditable workflows.