The Orchestrator

The clearest agent-native example is Beads, Steve Yegge’s lightweight issue tracker, which is also the data plane underneath Gas Town. Beads began as structured JSON — an ID, a description, a status, an assignee — tracked in Git alongside the project, and now keeps that state in Dolt, a version-controlled database.⁵ Its design answers a failure mode anyone who has run multiple agents has met: if you let agents manage their own to-do lists in text files, they will check off items they did not actually complete, and when challenged they will point at their own checked list and insist the work is done.⁶ Text files are self-reported. A task system that is a database — where status changes are structured events, work cannot be marked complete without meeting the conditions, and the ordering of steps is enforced rather than promised — gives agents something they cannot talk their way around. That is the same principle this book has been making since Chapter 6: guidance for the model, enforcement outside it.

The commercial incumbents are moving in the same direction. Atlassian, whose Jira began as a bug tracker and became the system of record for millions of teams, is now shipping agents as first-class users. The marketing line is telling — teams “plan work with AI, assign work to coding agents, monitor sessions, automate engineering loops” — because every one of those verbs used to be a human’s job.⁷ Jira’s twenty-year advantage is that it already holds the permissions, the context, and the audit trail; its bet is that the same system of record can govern agent work that used to govern human work.

Linear, the tracker many engineering teams adopted as the modern alternative to Jira, is competing on autonomy directly: it has spent 2026 turning agent integration into a first-class platform, so that third-party coding agents like Claude Code, Cursor, Devin, and Copilot appear as workspace teammates with their own profiles, and Linear’s own agent now sets up and runs coding sessions — writing, testing, and reviewing code — before handing the work back.⁸ Jira’s move is to make agents fit the system of record; Linear’s is to make the tracker fit the agents. An orchestrator evaluating this layer is watching the same question across all of them: is the board a system of record an agent cannot falsify, or a suggestion box an agent writes to?

One more form is emerging from the tools themselves: the plan-and-step board. Agent frameworks increasingly ship a structured plan — a set of steps with statuses, checkpoints, and provenance — that functions as agent Kanban — the same column-shaped board a human team would recognize, updated by the agent as it works. Hermes Agent, the runtime I work alongside, carries plans as first-class objects with steps, statuses, and restorable checkpoints; RiotPlan, a planning tool I built, builds the same pattern as a standalone service.⁹ The interesting property is the checkpoint: not just what is done, but the state of the work when it was done, so a run can be rewound or resumed. Human Kanban never needed that, because a human resumes from memory. An agent resumes from the record.