Chapter 10. Humans and the Swarm: The UX of Managing an Agent Fleet
In a single year — from August 2025 to June 2026 — every major vendor of coding converged on one interface model: "many agents, one screen," where the human sits in the 's seat rather than the interlocutor's. But this convergence has an uncomfortable flip side: the interfaces made _launching_ ten agents trivial, while _reviewing_ their results remained sequential human labor — and it is review, not model capability, that has become the 's limiting factor. Fleet UX is first and foremost the design of human attention, not the design of panels.
Converging on the Fleet Dashboard: Timeline and Shared Mechanics
The speed at which the industry arrived at a single pattern is itself telling. GitHub shipped the agents panel (August 19, 2025) — an overlay on every github.com page for launching and tracking Copilot coding tasks, then announced Mission Control as part of Agent HQ at GitHub Universe (October 28, 2025) and a standalone Copilot desktop app (June 2, 2026). OpenAI positions the Codex app as a "command center for agents" (announced ~February 2, 2026 — the date rests on secondary sources, the primary being unavailable). Cognition shipped Devin can now Manage Devins (March 19, 2026). Cursor 3 (April 2, 2026) made the Agents Window the primary interface, demoting the classic editor to an "add-on." Anthropic, in its Claude Code desktop redesign (April 14, 2026), put it plainly: "many things in flight, and you in the seat."
Under the hood they all share the same isolation mechanic: a git worktree per session. Claude Code places each session in its own worktree, the Copilot app runs "each session in its own isolated git worktree," Cursor and Codex do the same (sources); Devin isolates more coarsely — a separate VM per managed agent. Cheap isolation is exactly what technically lets one human run N agents on a single repository.
Claude Code, for its part, documents four modes of parallelism that differ in _who coordinates the work_: within a single session; the agent view (claude agents) — "one screen with all your sessions and the ones waiting for your input"; agent teams — a lead plus teammates with a shared task list (experimental, off by default); dynamic workflows — an orchestrator script that cross-checks results. The /batch skill splits one change into 5–30 worktree-isolated subagents, each of which opens its own PR. The question "who coordinates?" is the main design axis of fleet UX.
This is where the open part of the chapter ends. The rest opens with a free registration — no payment, just an email address.