Your Chief of Staff brings in the agents a job needs, but a new project often wants one of its own. Tap the faces beside the message box on Home and press New agent under your agents. The dialog asks one question: Claude Code, Codex or OpenCode. That is the whole form.

The name and picture are assigned for you, from a pool of human names you are not already using. Model, effort and machine start at their defaults, and you can change all three later on the agent’s screen. If you have not signed in a subscription for the tool you chose, you are sent to sign one in first (see Your subscriptions).

The new card reads Setting up — a minute or two, and it is ready in about that. If the machine cannot be built, the half-made agent is removed. Several agents can share one seat, and more seats run more agents at once (see Your subscriptions). Giving one its first job is in Talking to agents.

The faces beside the message box on Home open every agent of yours: the ones working first, then the most recently used. Each card is a glance at how that agent is doing.

• The context ring around its face shows how full its conversation memory is. It turns amber from 67% and red from 90%, when the agent should start a fresh chat.
• Its model and effort, and its machine shape.
• A state line: Now with the elapsed time while it works, Ready or Active 4m ago when it is not, and lines such as 2 proposals waiting on you or Session limit when something needs you.
• A one-sentence recap of its latest work, rewritten after every turn.

Tap a card to open the chat. An agent that is still setting up or out of session opens its own screen instead, and a paused agent with no subscription opens Your subscriptions.

When some of your agents have finished their work and nothing waits on them, a Reset finished agents button under the cards gives them all a fresh chat in one press.

Tap the name or face in the chat header to open the agent’s own screen. The screen shows the recap, the Thinks with and Machine rows, and the software card.

Tap Thinks with to open How ‹name› thinks. A choice saves the moment you tap it and begins on the next turn; a turn already running finishes as it started. The catalog depends on the tool.

• Claude Code: opus, the default, fable and sonnet. Effort: medium, high, xhigh, max.
• Codex: gpt-6.1-sol, the default, plus gpt-6-astra, gpt-6.1-terra, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna and gpt-5.5. Most offer medium through ultra; the luna models stop at max, and gpt-5.5 at xhigh.
• OpenCode: the OpenCode Go models, glm-5.3, the default, plus kimi-k3, deepseek-v4-pro, qwen3.8-max, mimo-v2.6-pro, glm-5.3-flash and deepseek-v4.1-flash. Each offers its own levels, and some have none.

The default effort is high wherever the model has it. The heavy levels, xhigh, max and ultra, are marked Much more compute: they think much longer and spend the subscription much faster. Worth it on a hard problem, expensive as an everyday setting.

Each agent has a dedicated virtual computer in your company’s cloud account, with a full shell, outbound internet and git. New agents are built on 2 dedicated cores and 4 GB. The other sizes are in Machine sizes.

To resize one agent, tap the Machine row on its screen. The picker offers dedicated and shared cores and marks the rung new agents get as standard. A size change never touches the agent’s files. A sleeping machine simply wakes on the new size. An awake one restarts for about a minute, after you confirm with Change size, and while the agent is working the change waits for the turn to finish.

Only the agent’s home folder survives a machine software update, so ask for anything that matters to be pushed to a repository. Resizing is yours to do; an agent never changes its own machine.

A machine sleeps about three quiet minutes after its last turn, and wakes in seconds when you send a message. The Machine row on the agent’s screen says which it is, and when asleep adds It wakes on your next message.

Sleep is safe. Every file and every unfinished piece of work is there when it wakes. A running background job keeps the machine awake, as do a turn in progress, setup, an open terminal and an update being installed.

A visitor can wake a machine too. A page the agent serves from its own address is reachable while it sleeps, but the first load takes about ten seconds. For hosting that never sleeps, see Sites and domains.

A workplace keeps all its agents in one list. Every agent reaches the same repositories, notes, knowledge pages, sites and Drive. To catch up after a day away, ask your Chief of Staff on Home.

To keep work apart, a client’s project or a second business for example, give it a workplace of its own. An agent stays in the workplace it was made in; see One workplace, one place.

Teammates’ agents appear under your own, grouped by the person who drives each one. Tapping one opens a read-only page; only its owner can message it or change how it runs.

Machines update themselves: each is brought up to the pinned version of Claude Code, Codex or OpenCode just before it goes to sleep. Most of the time there is nothing to do. The software card on the agent’s screen shows where it stands, and when a version is waiting you can press Update Claude Code (or Codex, or OpenCode) or Update machine software yourself. The last one asks first, because it restarts the machine for about a minute; files are kept. Both buttons are locked while the agent is working.

Delete this agent is at the bottom of the agent’s screen. You type the name to confirm. The machine and everything on it are erased for good, its private keys with it, and the chat stops answering. Anything never pushed to a repository or uploaded to Drive goes too. A retired agent, whose owner has left the workplace, has the same button.