An agent is three things at once: a name you write to, a machine that does the work, and a conversation that is the record of everything you asked and everything it answered. Every agent has all three, and none are shared with another agent.

The machine is a dedicated virtual computer, created in your own Fly account when the agent is made. Nobody else runs on it. It has a full shell, outbound internet, git, and the tools a developer would install, and the agent can serve one page from it at a stable address of its own.

An agent runs one of two coding tools, chosen when it is created and kept for life: Claude Code, which runs on your Claude subscription, or Codex, which runs on your ChatGPT subscription. The choice decides which subscription pays for its work and which models it can think with.

On its machine, only the agent’s home folder is durable. Everything else can be rebuilt when the machine software is updated, so the agent keeps its work in that folder and pushes anything that matters to a repository. The platform also keeps a manual on the machine, named CLAUDE.md for a Claude Code agent and AGENTS.md for a Codex agent, which tells the agent how the Workspace works. The agent’s screen shows which version of that manual it has.

A room is a team inside your workplace. A new workplace opens with two, named Room 1 and Room 2, and you can rename them and add more; an agent always belongs to a named room, so a room has to exist before an agent can be created in it. Agents, notes and secrets belong to a room. Repositories, knowledge pages, sites and Drive belong to the whole workplace, so every room’s agents reach the same code and the same pages.

A room is drawn by the faces of the agents in it, never by a folder icon. A room row shows up to three overlapping faces, then a count such as +6, and under them a line like Alex, Ben and 6 more or No agents yet. An empty room shows a dashed ring. On Home the rooms are chips along the top of the room half, and Manage rooms opens the list, where you can drag rooms into the order you want.

To move an agent to another room, open the agent’s screen and tap its Room row. The picker is titled Move ‹name› and says what the move changes: what the agent can reach. Its machine and its chat stay put, and the move lands at the agent’s next turn. A moved agent stops receiving the old room’s keys; see A key belongs to a room.

Teammates’ agents appear in a room too, in a compact strip grouped by the person who drives each one. Tapping a face opens a read-only page for that agent with its State, Latest work and Driven by rows. Only its owner can message it or change how it runs. For rooms themselves, read Workplace and people.

Open a room on Home and press New agent at the foot of the Agents section. With no room chosen the button reads Create a room first and does nothing else. If the workplace has no cloud connected yet, a step titled Connect Fly to create your first agent comes first; an admin connects it there, and a member is told to ask an admin.

The dialog asks one question: what the agent should run. Choose Claude Code or Codex. That is the whole form. The name is assigned automatically from a pool of human names, skipping any you already use, and the picture follows from the name. Model and effort start at their defaults, the machine at the size an admin has set for new agents, and you can change all three afterwards on the agent’s screen. If the platform restarts while an agent is still being set up, the setup finishes on its own; there is no need to press New agent again.

An agent needs a subscription to run on. If you have not signed in a Claude or ChatGPT subscription for the tool you chose, the Workspace sends you to your seats first, with a line such as A Claude Code agent runs on your Claude subscription — sign it in here first, then create it. See Your cloud and subscriptions.

Then the agent’s card appears with a three-step meter and the line Setting up — a minute or two. It is ready in about that. If the machine cannot be built, the half-made agent is removed and you are told nothing was left behind or billed. There is no per-person limit on how many agents you create.

The Agents section of a room lists every agent of yours in it, most recently used first, with a count line such as 3 total · 1 active · 2 idle. Each card shows, top to bottom:

• Its face inside a context ring, and its name. The ring shows how full the agent’s conversation memory is; its label reads Context window 42% full. It turns amber from 67% and red from 90%, when the agent is nearly full and should start a fresh chat. The Workspace Agent’s ring turns earlier, at 40% and 60%.
• Model and effort, such as gpt-6-sol · high. The tool is left off because the model implies it.
• The machine shape, such as 2 dedicated cores · 4 GB.
• A state line: Now with a live dot, the elapsed time and the current step while it works; Active 4m ago or Ready when it is not; 2 proposals waiting on you when it has proposed something; Setting up — a minute or two; Session limit and later Ready to resume when its subscription window is spent; Paused — sign in Claude, Paused — choose a subscription or Paused — Claude subscription needs attention (or ChatGPT) when it has nothing to run on; Retired — it can no longer run.
• A pill reading 2 background jobs while any are running. Tapping it opens the jobs screen.
• A one-sentence recap of its latest work, rewritten after every turn and at checkpoints during a long one.
• Thumbnails of pictures from its newest message, and the message that started the turn now running.

Tapping a card opens the chat. An agent that is still setting up, out of session or retired opens its own screen instead, and a paused agent with no subscription opens your seats. To reach an agent’s own screen from its chat, tap its name or face in the header.

Every agent thinks with one model at one effort level, and you can change both at any time. Tap the model line under the agent’s name, in the chat header or on its screen, or the Thinks with row. The picker is titled How ‹name› thinks, with a Model section and an Effort section. The current choice wears a current badge, and a choice saves the moment you tap it.

Changes begin on the next turn. A turn already running finishes with the settings it started with.

The catalog depends on the tool the agent runs:

• Claude Code: opus, the default, and fable. Effort: low, medium, high, xhigh, max.
• Codex: gpt-6-sol, the default, plus gpt-6-astra, gpt-6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna and gpt-5.5. Most offer low through ultra; the luna models stop at max, and gpt-5.5 at xhigh.

The default effort is high for both. The heavy levels, xhigh, max and ultra, are marked Much more compute in the picker, and the note under the card says why: they think for much longer before answering and spend the subscription’s usage much faster, so a few long turns can drain a window. Worth it on a hard problem, expensive as an everyday setting.

Each agent’s machine has a size: how many processor cores, whether they are dedicated or shared, and how much memory. New agents are built on 2 dedicated cores and 4 GB unless an admin has changed the workplace’s default under the cloud screen’s New agents row, which never touches a machine already running.

To resize one agent, tap the Machine row on its screen or the machine line in its header. The screen is titled ‹name›’s machine and offers two ladders:

• Dedicated cores: 1, 2, 4 or 8 cores of its own, with 2, 4, 8 or 16 GB. Never shared with anyone else.
• Shared cores: 1, 2 or 4 shared cores, with 2, 4 or 8 GB. Far cheaper, and throttled hard under sustained load; fine for an agent that mostly waits.

Each rung shows an approximate cost per awake hour, and the rung new agents are built on is marked standard. A machine costs processor and memory only while it is awake; asleep, only its disk. Those figures are estimates: Fly’s dashboard is the authority for what a machine actually costs, and Fly’s pricing page has the detail.

A size change never touches the agent’s files. If the machine is asleep, the change is free and interrupts nobody: it simply wakes up on the new size. If it is awake, the Workspace asks Restart ‹name›’s machine? and warns that anything running on it stops for about a minute; confirm with Change size. While the agent is working, the size can change as soon as the turn finishes.

Memory is a wall, not a slope. There is no spare memory to spill into, so work that outgrows the machine is stopped rather than slowed, and the fix is one rung up the ladder. Agents never resize their own machines; that is yours to do.

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, for example Awake · Amsterdam · 2 dedicated cores · 4 GB, and when it is asleep adds It wakes on your next message.

Sleep is safe. The disk is kept, so every file, every unfinished piece of work and the agent’s own notes are there when it wakes. What is not kept is a plain running process: anything the agent merely started in the background stops with the turn. Work that must continue runs as a background job, which is one of the things that keeps a machine awake. The others are a turn in progress, an agent still being set up, an open terminal session and a software update being installed.

A visitor can wake a machine too. Anything the agent serves at its stable address is reachable while it sleeps, but the first load of a sleeping page takes about ten seconds, and the page only comes back if the agent set it up to restart on boot. See Sites and domains for hosting that never sleeps.

The platform pins one exact version of Claude Code and one of Codex for every machine. Machines update themselves: an existing machine is brought up to the pinned version just before it goes to sleep, and a new machine is built with it already installed. Most of the time there is nothing for you to do.

The version card on the agent’s screen shows where it stands. The tool row reads ‹installed› · up to date or names the version available, and Machine software reads Up to date or Update available — it restarts the machine, about a minute. Before the first read it says the versions get checked when the machine next goes quiet.

You can update by hand. Update Claude Code or Update Codex installs the tool and reads Updating… this takes a few minutes while it runs. Update machine software asks first, because it restarts the machine: running apps stop for a minute and files are kept. When both are behind, the tool is included in the machine software update and has no separate button. Both are locked while the agent is working, and a second tap during an update just says it is already updating.

Besides the agents in your rooms, you have one Workspace Agent. It is private to you and reaches the whole workplace: it is outside every room, absent from room rosters and People counts, and nobody else in the workplace sees its conversation. Each member has their own.

It is a full agent with a durable machine of its own, built the first time you send it a message. It can work with files and code, use the web and run services, and it can coordinate your room agents when you ask: see their history, stop them, and send them work. When a room agent answers a request the Workspace Agent sent, the answer lands in the Workspace Agent’s conversation and starts a turn there.

On Home its card shows its face, name, model and effort, and its state: Ready, Working, or what it is waiting for. It reads Thinking while it works and shows its last reply beneath. Tap the card to open the conversation, or the microphone on it to talk instead. Its details screen has a Go to chat button, an Its room row for the notes it keeps for itself, and facts: State, Last request, Machine (or Not reserved yet before its first message) and Reach. The same version rows and the same Ask agent to reset chat button sit below, and its own picker, Workspace Agent model & effort, opens from the model line.

Under the Workspace Agent card, Home has a Workplace briefing: a private summary of the last seven days across all of your rooms, three to six bullet points, about a hundred words. It is written from your own conversations only, never from another person’s chat.

Nothing writes it on its own. Press Refresh and it is written for you; before the first one the card reads No briefing yet. Press Refresh to get one., and while it works, Writing your briefing…. A quiet week reads No activity in the last seven days.

The briefing has its own model control, next to the Refresh button, which opens Recap model & effort. Choices there are saved as you tap and used for your next briefing; they do not change how the Workspace Agent or any room agent thinks.

At the bottom of the agent’s screen is Delete this agent. The hint under it says what happens: deleting erases its machine and everything on it, for good, and its chat stops answering. Anything on that machine that was never pushed to a repository or uploaded to Drive goes with it, along with any private keys saved for that one agent.

The Workspace asks Delete ‹name›? and makes you type the name into Type the name to confirm before Delete will press. The card then reads Erasing its machine… until it is gone, and you are told the agent and its machine are gone. This cannot be undone.

A retired agent, one that can no longer run because its owner is no longer in the workplace, has the same button; deleting removes it from the list for good.