Fatih Arslan, writing on his personal blog on September 11, 2026, describes managing a fleet of agents through plain Markdown plans stored in a dedicated Git repository. A /plan-init skill creates drafts, next, open, done and discarded folders plus a README index, and skills named /plan-add, /plan-write, /plan-dispatch, /plan-sync and /plan-status move each plan through those states.
The central idea is that a versioned folder of human-readable plans can serve as the control surface for many agents. The mechanism is a division of labour: /plan-write uses a higher reasoning model to prove a cause and name the change, so /plan-dispatch can hand a complete plan to a fast coding agent that opens a pull request. Because the plans sit in Git, agents already know how to work with them, and history comes for free.
The timing matters because Arslan now runs these plans through coordinators, the long-running agents Cursor introduced with Projects, which subscribe to pull requests, Slack channels or schedules. He commits coordinator notes to the same repository, which mattered when a platform bug left a session stuck in an error loop after roughly 460 messages. Teams adopting coordinators should keep durable plan state outside any single agent session.
The account describes one practitioner's workflow and does not establish how it compares with other orchestration setups or larger fleets. Other Projects users publishing similar plan layouts with merge rates or failure counts would show whether the pattern generalises.