Multi-agent supervision: the fleet view arrives

Every agent-UX pattern on this site — approval gates, streaming progress, interrupts — was designed around one human watching one agent. That assumption broke sometime in the first half of 2026. Developers stopped running a single coding agent and started running several at once: one fixing a bug, one reviewing a PR, one investigating a flaky test. The interface problem changed from "how does a human supervise an agent" to "how does a human supervise a queue of them" — and three separate products shipped an answer within about six weeks of each other.

The list view: Claude Code's Agent View

Anthropic shipped Agent View as a research preview on May 11, 2026, launched with claude agents and requiring CLI v2.1.139 or later. It's a terminal table, one row per background session, and the row is doing real supervisory work rather than just listing processes.

Each row carries a status icon — animated while Claude is actively working, yellow when it's "needs input" (an approval decision, a question, a sandbox or MCP permission prompt, or a command that needs an attached terminal), dimmed when idle, green on success, red on failure. A separate glyph tracks whether the underlying process is still alive or has exited and can be resumed. When a session is stuck on "needs input," the row shows the literal question being asked, so a human can triage without opening anything — pressing Space opens a peek panel to answer in place. Rows also carry a one-line summary regenerated by a small model roughly every 15 seconds, and a colored pull-request badge once a session opens one. Each dispatched session runs in its own worktree automatically, so parallel agents can't collide on the same files. Anthropic is explicit that this is separate from subagents (spawned inside a single conversation, not independent jobs) and agent teams (which add cross-session messaging) — Agent View's job is specifically the queue-of-independent-jobs case (Agent View docs).

The board view: Devin Desktop's Agent Command Center

Cognition took a different visual metaphor for the same problem. On June 2, 2026 it rebranded Windsurf into Devin Desktop, replacing the editor-first default view with a Kanban board — the Agent Command Center — as the landing surface. Columns sort agents by status (in progress, blocked, ready for review), and the board mixes local Devin Local agents, cloud Devin agents, and third-party agents connected over the Agent Client Protocol, including Codex, Claude Agent, and OpenCode sessions running side by side (Windsurf is now Devin Desktop; Devin Desktop FAQ). A companion feature called Spaces groups agents by project, sharing context — PRs, files, prior tasks — across sessions instead of treating each one as a throwaway conversation. The transition shipped as an ordinary update: existing Windsurf settings, keybindings, and MCP connections carried over with no migration step, and Cascade, the old local agent, was deprecated on July 1, 2026 in favor of Devin Local.

The Kanban choice is worth noticing next to Claude Code's flat list: a board implies workflow stages an agent moves through (a claim this site can't yet verify against a shipped roadmap feature), while a list implies independent, unrelated jobs. Same underlying problem, different bet on how much structure to impose on it.

The grid view: Cursor's Background Agents panel

Cursor took a third shape. Background Agents run as remote jobs in isolated cloud VMs — Cursor cloned the repo, worked on a separate branch, and (since a February 2026 upgrade) gave each agent its own desktop environment and browser so it could interact with UI elements, not just edit files. A sidebar panel lists every running agent with its current step, elapsed time, and a live log, and Cursor's Agent Tabs let several of those conversations sit side by side in a grid rather than a single stacked list. Reporting on Cursor's usage puts typical parallel counts at up to eight background agents running at once (Cursor Background Agents guide; Cursor's agent mode overhaul).

What all three converged on

Strip the visual metaphor away — list, board, grid — and the three products agree on the same primitives:

None of them solved supervision by making the dashboard smarter than the human. All three solved it by making triage cheap: a glance tells you which of N agents is stuck, and stuck always means the same thing — it's waiting on a decision only a person can make, surfaced as specifically as the underlying approval-gate design allows. That's the same principle this site's guide to streaming progress argued for a single agent — narrate at decision grain, not action grain — just applied across a queue instead of one conversation.

What's still unsolved

None of the three products have shipped a real answer to the harder question: what happens when two agents in the same fleet touch the same resource — the same file outside their worktree, the same external API, the same customer record? Isolation (worktrees, VMs) sidesteps collision on disk, but an agent fixing a bug and an agent refactoring the same module can still produce conflicting PRs that only a human resolves at merge time, and none of these dashboards show that risk before it happens. Nor is there a shared standard for the "needs input" signal itself — Claude Code's sandbox/MCP/question triad, Devin's "blocked" column, and Cursor's log-based cues are three incompatible vocabularies, which means a human context-switching between products relearns triage each time.

The fleet view is genuinely new interface work, not a repaint of single-agent UX — but it inherited the same open problem this site keeps finding at the human layer: approval semantics fragment across every product that builds them, and multiplying the number of agents just multiplies how often that fragmentation costs someone attention. Worth tracking as more products ship their own version of this dashboard.