Introducing the 'Chief of Staff' Orchestration Pattern to Solve Claude Code Long-Running Issues

asyncdot.com points out that the causes of long-running execution failures in Claude Code are not a lack of model capability, but rather ephemeral context loss and unreliable self-report, and releases the 'Chief of Staff' pattern to solve this. This pattern adopts an Orchestrator-Worker structure where a Coordinator (long session), which does not perform implementation directly, handles assignment and verification, while a Worker (short session) performs the implementation.

Core Operational Loop and Verification Principles

The operational loop proceeds in the order of PULL → READ → RED GATE → DELEGATE → PROVE → OBSERVE → GATE → SHIP, ensuring 1:1 traceability between Git history and the board through the 1 item 1 commit principle. In particular, the RED GATE stage confirms whether the verifier fails first to prevent unnecessary code review for tasks that have already passed.

During verification, the agent's report is evidence, not instruction, so the Exit code is used as the authoritative standard and diff hunk is directly checked. Additionally, values are read back from artifacts rather than variables for verification, defending against flaws such as Vacuous assertion by distinguishing between '0 items' and 'execution failure'.

State Management and Communication

Sessions are treated as disposable, and state is stored in a Durable state store (e.g., Plan Desk). Since direct messaging between sessions is unreliable, essential information is communicated via Durable channels such as Committed files or Board cards ("Message is a nudge, file is a contract").

Application Conditions and Tools

This pattern is suitable for multi-session parallel work or long-term projects, but is unsuitable for single changes or cases with high overhead. In terms of tools, cmux or Plan Desk are not mandatory, but the ability to create isolated agent sessions and an agent-accessible Durable task store (API/MCP) are core requirements. In a Mixed-model configuration, having a model family different from the implementation model perform verification can capture systematic blind spots.