Why Agents¶
Keep your editor and your coding agent. Agents is where you run several of them at once, see which one needs you, let them work together and on their own, and answer them from your phone.
It is not an editor¶
An editor with an agent in it is built for one task at a time: you ask, you watch, you accept. Agents is built for five agents on five tasks while you are in a meeting. It has no editor of its own. Open the same folder in yours whenever you like.
It runs the agents you already have¶
Agents starts the coding tools you already use (Claude, Codex, Gemini, Antigravity, Grok, Copilot and Cursor), each signed in with your own account and plan. The ones it can install, it installs in a folder of its own. It brings no model and no bill of its own, and it is open source under the MIT licence. You can choose a different runtime, model and permission mode for each agent.
They do not all do the same things; Runtimes lists each.
It shows which agent needs you¶
Every agent sits in one group: Needs you, Waiting, Working, Done or Paused. Needs you holds questions, unresolved blocks, endings nobody accounted for and unexpected stops; a finished turn you haven't read keeps an unread mark wherever it is; Waiting carries on by itself. When an agent ends its turn, it says in one sentence how it went. You can read the list and know what to open without reading any conversation. See Statuses and groups.
You see the work, not just the chat¶
When an agent writes a document for you, such as a plan, a spec or a review, it opens beside the conversation and fills in as the agent writes. Each change is marked, and the page moves to it. You can type on the page yourself while it is being written, on the Mac, the iPhone or the iPad. Diagrams arrive as pictures in the page, not as text to imagine. See Follow a live document.
When it changes code, the Changes pane lists every file it touched, with each edit coloured down to the word. See Read an agent's changes.
Agents that work together, and on their own¶
Agents take turns with what they share, wait for things to happen, and start when something happens, without you.
Here is one afternoon. An agent finishes a change, and a workflow, a Markdown file kept in the repository, starts another agent to test it. The tester needs the iOS simulator to reproduce a failure, but another agent is using it. So it asks for a lease, gets in line and ends its turn. It spends nothing while it waits, and it is started again when the simulator is free.
It reproduces the failure, fixes it and commits. Then it
waits for main to move so it can rebase. Its turn
ends again, and nothing polls. When main moves, it is woken and told so. It rebases and
finishes, and a release agent that was waiting for it wakes up and starts on the release
notes.
You were not asked for anything. The list shows each step as it happens.
Each part can be used on its own:
- Leases. Agents take turns with anything only one of them should use at a time: a simulator, a browser, the screen, or anything an agent names. The Mac's Resources page shows who holds what, and you can end a lease. See Leases on shared resources.
- Events and waiting. An agent can wait for another agent finishing, a branch moving, the Mac waking, or you coming back. The Mac keeps a log of events, and the phone and iPad can read it. See Events.
- Workflows. A workflow starts an agent on a schedule, such as every weekday at nine, or on an event, such as an agent finishing or a branch moving. It is a file beside your code, reviewed like code. See Set up a workflow.
- Agents that manage agents. An agent can start others (three running and five kept at once, unless you change it per project), brief them, wait for them and stop them. Each can work in a worktree of its own, so they never edit the same files. See Tools the app gives agents.
Claude and Cursor may ask your permission before an agent uses some of these tools.
They keep working without you¶
Your agents belong to a helper that runs in the background, not to the window. Close the window or quit the app, and they carry on. They can still ask you questions, and the questions wait. While a turn is in progress, the Mac does not go to sleep on its own. If the Mac restarts, the agents that were working pick up where they stopped once you log in again. See The window and the host.
On your iPhone or iPad you see the same list. You can read a conversation as it happens, answer questions, send a prompt or a picture, start and stop agents, and open their files and a terminal. When an agent needs you and you are away from the Mac, the device you used last gets a notification. See How the phone and iPad reach your agents.
Your machines, one figure¶
The work happens on your Mac, or on a Linux server of your own. There is no service of ours in between. See Add a server.
What every agent spends, on every runtime and server, adds up to one figure for today. You can set a limit per agent and per day. See Settings.
Compared with what you use now¶
An agent in your editor (Cursor, VS Code with Copilot, Windsurf, Zed). You keep your editor. Agents adds many agents at once, live documents, watched pull requests and everything above. You give up editing and chatting in one window.
An agent in a terminal (Claude Code, Grok, Cursor's agent). You keep the same tool, because Agents runs it, and adds a window over all your sessions. In the conversations Agents starts, the tool's own scheduling, notifications and reports go through the app instead; from Terminal it is unchanged. Its sub-agents are left alone — the agent runs them itself, and Claude's and Codex's are drawn over the prompt so you can watch them. See Why agents' own tools are taken away.
An agent in someone's cloud (Copilot's coding agent, Codex in the cloud, Claude Code on the web). The work stays on your machines, with your own tools, files and simulators. You give up having someone else's machine work while yours is off.
Another app that runs several agents. Compare on the runtimes you already use, leases, events, workflows, the phone and servers.
When something else is better¶
- You want one assistant beside the file you are editing. Your editor's agent is closer to hand.
- You are not on a Mac. Agents needs a Mac. Servers and phones hang off one.
- You want to download and go. Today Agents is built from source in Xcode.
- You want to reach your agents with the Mac asleep. Agents on your Mac stop being reachable while it sleeps, and so does everything else if the control plane runs on it. To work with the Mac asleep, put the work on a server and run the control plane as copies elsewhere (see The control plane).
- Your machine should be off while the work happens. A cloud agent does that; Agents does not.