One screen for all my agents: Claude Code's agent view

Running one Claude Code session is easy. Running six is where it used to fall apart: a terminal tab per agent, alt-tabbing around to check which ones finished, which ones are stuck waiting for a permission prompt, and which one quietly asked me a question twenty minutes ago.

Claude Code’s agent view fixes exactly that. Run claude agents and you get a single screen with every background session as a row: what it’s doing, a summary of its latest activity, and whether it needs you. It shipped as a research preview in May 2026 (v2.1.139+), and it has quickly become my default way of opening Claude Code.

The dispatch board

The mental model shift is the interesting part. Instead of “open a terminal, start a session, babysit it”, agent view works like a dispatch board. There’s an input at the bottom; you type a task, hit Enter, and it becomes a new background session running on its own. Type another one. And another. Each row keeps working with no terminal attached — a supervisor process runs them, so you can close the window, even close your shell, and the work keeps going.

The list is grouped by what matters: sessions that need input float to the top, then ready for review (the ones with an open pull request), then working, then completed. So triage is a glance: anything at the top of the list is waiting on me; everything else is making progress without me.

From a row you can peek at a session to read its latest output and reply inline, or attach to drop into the full conversation — at which point it’s a normal Claude Code session, every command and shortcut works. Press ← and you’re back at the board.

Getting sessions in there

Three ways, and they compose nicely:

• From the board: type the prompt in the dispatch input. @subagent names work, and /skill commands dispatch as the first prompt.
• From inside a session: run /bg and the conversation you’re in moves to the background and shows up as a row. Great for “this is going to take a while, I’ll check back”.
• From the shell: claude --bg "investigate the flaky test" starts a session that goes straight to the background. You can name it (--name), pin a model, or even run a specific subagent as the main agent.

Parallel sessions in the same git repo don’t trample each other: before editing files, each one moves into its own isolated worktree under .claude/worktrees/, so they all read the same checkout but write separately. One thing to keep in mind: each session burns through your subscription quota independently, so a big fleet adds up roughly linearly.

Why this changes the day

Before agent view, the number of agents I ran in parallel was capped by my tolerance for tab juggling — in practice, two or three. Now the cap is how much independent work I can articulate. A bug fix, a PR review, and a flaky-test investigation are three rows, not three terminals. I dispatch them in the morning, work on whatever needs a human, and check the board the way you’d check a kitchen ticket rail.

It also pairs beautifully with prompting by voice: since v2.1.145 the dispatch input supports push-to-talk dictation, so “describe task, release key, Enter” is the whole loop. Saying what I want, into a board that fans it out to as many agents as needed, feels much closer to delegating to a team than to using a code editor.

It’s still a research preview, so the interface and shortcuts may move around. But the direction is clear: the unit of work isn’t the session anymore — it’s the fleet, and the fleet finally has a control room.