Your agent team

Every agent has a role, and the role writes its opening briefing — what to do, what to check first, what to hand back to you. It is filled in for you in the hire dialog, where you can rewrite as much of it as you like before or after hiring. The role stays a label besides: it is how the team routes work to each other, and it is the face at the desk.

How many agents you can have is your plan, and the room is the plan. Free seats one and gives you the desk— a one-desk room with a spare beside it. Starter seats three in the studio, and Pro and Business open the office, which has eight. Nobody is ever sold a colleague with nowhere to sit, and an upgrade is somewhere to put them rather than a number going up. Let one go and the seat comes straight back.

The panel on the right holds two conversations, because they are two different relationships. Assistant is the agent that builds your product with you. Teamis the agents you hired to run it — they answer with that project's data, knowledge and permissions.

Ask the team and they decide who takes it; the reply says who, and why, so a wrong hand-off is something you can see and fix rather than wonder about. Or pick someone by name from the selector. Either way the answer is attributed, and it never exceeds what you allowed that agent: one in Suggest replies with a draft, clearly marked nothing has been done yet. Asking directly does not raise anyone's permissions.

Threads are kept. Reopen one from the history icon and it still names who answered — even if you have since deleted that agent, because the record is of what happened.

Your agents have names, roles and a voice you write, and the chat panel on your site shows whoever is answering — Ada, support. What it also shows, always, is an AI assistant marker in the header and one line at the top of every conversation: you are chatting with an AI assistant, not a person.

That line is not a setting. You can write the greeting, the panel title, the subtitle and the placeholder — but not that, and it cannot be switched off. Since 2 August 2026, Article 50 of the EU AI Act requires anyone providing an AI that talks to people to make sure the person knows, from the first message. The obligation is ours as the provider, and it applies wherever your visitor is, so we carry it for you rather than leaving it in your copy to remember.

Practically: nothing to configure, nothing to switch on, and one less thing to get wrong. Everything else about how your agents present themselves is still yours.

A routine is a standing appointment — “every morning at 9, check for stuck orders and summarise”. Set them on the project's Teamtab, inside the agent's editor, from presets or your own cron (at most every 10 minutes — each run is a real agent run).

A schedule is initiative, never permission: the routine fires through the same rules as everything else the agent does. In Suggest a routine only produces drafts; in Act it works within its caps, and sends still queue for your approval unless allow-listed. The builder can set routines up when you ask — but no agent can ever schedule itself.

The event-driven twin of routines: “when a form is submitted, evaluate the lead”. Set them in the same place — the agent's editor on the project's Team tab — for form submissions, new table rows, or failed workflows.

Same rules as everything else: the firing runs under the agent's own permission, and no agent can set a watch on its own. One guard is specific to watchers: each fires at most once per 10 minutes, which also breaks the loop where an agent's own action re-triggers its watcher.

One email a day: what the team did in the last 24 hours, what waits on you (drafts to review, actions to approve, questions to answer), and what broke (failing routines, watchers, workflows). Turn it on under Settings → Daily standup and pick the hour — it sends at or after that hour in your timezone, at most once a day. Days with nothing to report send nothing.

Every line is a plain count read straight from your workspace — no agent writes this email, so it can't embellish a success or soften a failure. You can also ask the builder to set it up (“send me a standup every morning at 8”); like all standing initiative, only you can switch it on.

An agent that hits a question it cannot answer from its briefing, knowledge or data doesn't guess — it asks you. Questions appear at the top of The room (and in your notifications): who asked, on which project, and what prompted it.

Answering is the teaching moment, and you choose who gets taught: the asking agent (default — the same question never comes back), the whole workspace(the answer lands in the handbook every agent reads — right when the answer is company policy, not one agent's detail), or nobody (one-offs that must not become policy). You can also answer from the builder chat (“tell Mühür the wholesale price is 40/kg”); asking needs no permission, but only you can answer — an agent can never close its own question.

The room on your dashboard is not a picture. Hover anything and it names itself; click an agent and their card opens — who they are, what they are doing right now (drawn from the same live state as the room, so the card can never disagree with the figure), with a composer that hands your words straight to the Team chat addressed to them. Click your own desk to jump to whatever is waiting on you; click the hire desk to grow the roster.

When something truly needs you, the room says so out loud: an agent waiting on an approval, stuck, or holding an unanswered question carries a speech bubble over its head — click it and you land where the thing is dealt with. Quiet room, quiet page: bubbles only ever show what is true.

Whoever needs you comes to your desk. An agent with an approval pending, and an agent whose run has stopped, both leave their desk and stand at yours facing it — a stopped run needs a person as much as a decision does. Several at once form a line around the desk, one to a tile, whoever has waited longest at the front, so you can see how much is queued up without opening anything. Their cards say which of the two each one wants.

You can also rearrange the room by hand: drag a figureonto another desk to reseat them. Drop on an empty desk and they move; drop on a colleague's desk and the two swap — nobody is ever left standing. The new seating is saved to your workspace (the same layout the Settings office editor shows), and the agents walk to their new desks in front of you. And yes — the arcade in the corner does something. Click it.

The room is always on screen — it is not a tab you switch away from. The rail under it changes what you are reading; the office stays. The expand button beside the zoom controls fills the screen with it (Escape leaves), the whiteboard opens the handbook, the filing cabinet opens your audit log in a window over the room, and the printer goes to the pages waiting for you.

The furniture is yours too. The paintbrush above the room turns on decorate mode: plants, bean bags, stools, lamps, side tables — even the arcade — can be carried to a new tile, and everything hung on a wall slides along its wall. The landing tile glows green when the spot works and red when it doesn't: the room refuses a drop on someone's seat, on another piece, or anywhere that would wall your team away from your desk — agents walk around furniture, so the walkways are real. A rugis the exception: it lies flat and is walked on, so it slides under the furniture and even under a chair — the only thing it cannot do is hang off the edge of the room. Moves save to your workspace, agents mid-walk reroute immediately, and pieces you never touch follow the room's own layout. You can also add furniture — the +button opens a catalogue of plants, seating and useful things; pick one and click a tile to put it down. Your builder agent can decorate on request too (“put a plant next to my desk”) — it obeys the same rules your drag does, on the wall as well as the floor.

The keyboard is quicker. With a piece selected, the arrow keys move it one tile (a hung piece one step along its wall), R turns it, Delete takes it out of the room, and Esc lets go. Ctrl-click picks several pieces at once — turn, paint, move or remove them together, in one change. When a spot is refused the room says why: something is already there, somebody stands there, or it would block the walk to your desk.

The wall is a surface too. Pictures, posters, boards, clocks, garlands and neon words hang on either back wall, and in decorate mode the cyan patch on the wall is the piece: the stretch it covers and the height it hangs at, painted flat on the wall itself. Slide a picture along its wall the way you drag a plant across the floor — and carry it round the corner onto the other wall, which turns to face you as it lands. Framed art and posters also turn — the turn button stands them on their end, which takes one cell of wall instead of two, so a portrait picture fits a gap a landscape one cannot. Two pieces can share a stretch of wall as long as they hang at different heights: a poster under a garland is fine, a poster over a whiteboard is not.

Your office can be a window as well as a workplace. In Settings → Office appearance → Public office link, one switch mints a URL anyone can open — no account needed — showing your room live: your agents at their desks, walking, working, the night palette after dark. It's the same renderer you watch, pointed outward.

What crosses that link is presence only. Task texts, approval contents, questions, customer data and internal ids never leave the workspace — the server strips the picture down to figures, desk names and activity before anything is sent. Turning the link off kills it immediately; turning it back on mints a fresh URL, so a link that reached the wrong audience is always revocable. Your builder agent can manage it too (“make my office public”) — enabling counts as an outward-facing action, so it asks first like any external tool. Pasted into a chat or a feed, the link unfurls into a picture card of the office at night — the room does its own advertising.

The Taskstab under the room is standing work — things you want done that outlive a chat message. Write a task, assign it to an agent on that project (or leave it in the backlog), and the agent carries its open tasks into every run: a routine like “work through your board every morning” actually functions. When the work is done, the agent hands the task in for review with a written report.

“Run now” starts it immediately. Open a task and hand it straight to the agent it is assigned to instead of waiting for their next routine. They work it under their own permissions — asking is not a promotion, so a draft-only agent still only drafts and an ask-mode agent's sends still queue for you — and the task moves to Doing while they work, then lands in Review with their report on the card. If the run fails, the reason is written onto the task rather than lost. You can ask for it in the sidebar too: “have Nova start the refund task”.

Only you mark work done— the same rule as drafts, applied to work items. An agent moves its own tasks forward (to doing, then review with a report), never a teammate's, and never to done; assigning tasks and setting them running are yours alone (the builder can do both when you ask, no agent can on its own — an agent that could push work onto a more privileged teammate and start it would be routing around its own limits). Tasks waiting in review show up in your daily standup.

An agent hired into more than one project has a board in each, and only ever sees the board of the project it is working in.

When a project runs more than one agent, each knows who else is on the team. A visitor who asks the support agent a pricing question gets handed to sales mid-conversation: the widget shows a quiet line — “Bera brought in Deal — pricing questions are Deal's area” — the header switches to the teammate, and the thread continues where it was. No repeated questions, no fresh greeting.

And the room shows it.Every hop is recorded now — who passed what to whom, and when — so the handoff is a fact about your team rather than something only the visitor ever saw. For the few minutes after one, the agent that handed the visitor on walks to the receiving agent's desk and tells them, mid-sentence. The agent that took over stays where it is: it is answering the visitor, and work keeps an agent at its desk.

The handoff is bounded by construction: an agent can only point at a teammate that exists, is enabled, and works on the same project; it can never transfer to itself; and the teammate who receives the visitor cannot transfer again on the same message — one hop, then someone answers. The visitor can always switch agents manually from the widget's picker too; a manual switch starts a fresh thread, a handoff never does.

Agents remember. When a run teaches one something worth keeping (“this customer prefers email”, “that import chokes on extra columns”) it saves a short note, and carries its notes into every future run — below your knowledge and the handbook, which always outrank them. Memory holds 50 notes; when it's full the agent must forget something first, so nothing is ever silently dropped.

There is no private state. The brain icon on an agent's card shows every note it has, you can delete any of them, and remembering or forgetting shows in the activity log like any other action. An agent can only touch its own notes — never a teammate's.

Every agent's card under Assignmentscarries its last 30 days in numbers: runs and failures, the share of its drafts you kept versus dismissed, questions asked, and anything stuck waiting for your approval. The same numbers are one question away in chat (“how is Mühür doing this month?”).

Nothing grades your agents — the acceptance rate is literally your own clicks on their drafts, and an agent whose work was never reviewed shows no judgment yet, not 0%. Asking questions is counted but never scored against an agent: asking instead of guessing is what you want. What to do about a bad month — retrain the briefing, narrow permissions, pause — stays your call.

The Handbooktab under the room holds workspace-wide facts and policy — shipping times, price floors, tone. Every agent on every project carries it into every run, on top of its own knowledge, and it outranks an agent's own notes when they conflict. Write the things you find yourself repeating, the way you'd tell a new hire.

Only you write it. The builder can add entries when you ask (“add to the handbook: we never discount past 20%”), but no agent can edit the handbook on its own initiative — an agent that could teach the whole team could mislead the whole team.

The chat window on a site and the agent behind it are two different things, and they are configured in two different places. The widgetbelongs to the site: its colour, bubble icon, corner, panel title and placeholder all live under that project's Team → Widget sub-tab, alongside — but separate from — the Agents sub-tab listing who works there. The agent belongs to your workspace and can staff up to three of these doorways at once.

This is why the look is set per project rather than per agent: one support agent working across three sites cannot have three colours, and switching teammates in the widget's picker should not restyle the panel mid-conversation.

The inbox is a second doorway, staffed the same way. On the project's Email page, choose who answers arriving mail: a named agent, the whole team (they route each message), or nobody — the default, where mail simply waits for you. Replies follow the agent's own permission exactly as everywhere else: Suggest drafts, Ask queues the send for your approval, and even Act queues sends unless you allow-listed them. Turning this on is yours alone — the builder can do it when you ask, but no agent can ever staff an inbox on its own.

The chat widget on your live site automatically embeds whenever any agent on the project is enabled. When you have a single agent, visitors just chat with it. When you have two or more live agents, the widget shows a small Team ▾switcher in its header — visitors can pick which teammate to talk to, and each switch starts a fresh thread with that agent's own greeting and persona.

Everything above is also available to the builder agent in the chat panel — so you can just say “add a sales agent that qualifies leads and captures their email.” Under the hood it uses the same tools you'd use by hand:

Whatever you can do from the Agents tab, the agent can do for you — the manual UI and the agent tools stay in lockstep.

Your agents don't only answer visitors — a workflow can wake one to do a job. Add an Ask agent action to any workflow (Automation → Workflows), pick the role that should handle it, and write the task. When the workflow fires — on a schedule, an inbound email, a database change — the agent produces its output.

The draft text is also available to later workflow nodes as {{agent_reply}}. A pending draft raises a notification so you know it's waiting. The builder agent can read your drafts too (list_agent_drafts) and build these workflows for you (“when a support email comes in, have the support agent draft a reply”).

By default every agent is Suggest— it only drafts. When you're ready for it to actually dothe work, raise its autonomy on the agent's Autonomy panel (the Agents tab, under the editor). There are three levels:

Guardrails, always on: agents can never delete or destroy anything autonomously — those actions are simply unavailable to them. Every agent has caps (actions per run and per day; over-cap work is queued for approval, never dropped). Two kill switches stop everything instantly: a per-agent pause on the Autonomy panel and a workspace-wide Pause all agents on the room. Every action — run, blocked, queued, or executed — is recorded on the Activity tab.

You raise autonomy; a trigger's content never can. An inbound email or webhook can give an agent a task, but only you (via the dashboard or the builder agent) can grant it more power.

If you have switched your phone on, an agent taking a call is drawn at its desk as on a call — in the room, on hover, and on its card. It is the strongest thing the office says about anybody, because it is the only state with a person waiting in real time: it outranks even work sitting on your approval. The moment the caller hangs up, the agent goes back to whatever it was doing.

Like everything else in the room, it comes from a row rather than an animation: a voice conversation that has not ended yet. A call whose provider dies mid-conversation ages out rather than leaving somebody holding a phone forever.

Autonomy is how far an agent goes on its own. Scope is what it can reach at all — and they are different questions. A support agent set to act should still not be able to delete a project.

Open What this agent can touch on the Autonomy panel. Every tool is listed under what it does — reads, writes, reaches out, removes — and you tick the ones this agent needs. Leave it alone and the agent can use everything, which is the default and is right for the first agent in a new workspace.

A tool you leave out is not refused mid-task: it is invisible. The agent is never offered it, so it does not try, and it never tells your customer it was blocked from doing something. That is the difference between a scope and a wall.

Your agents know their own limits.Every autonomous run starts with the agent being told, in plain words, what it may actually do: how many acting steps it has this run, how much of today's budget is left, which tools it is restricted to, which saved connections it can call, and which sends — if any — go out without your approval. So when you ask one of your agents “what can you do here?” the answer is read off the same numbers the platform enforces, not guessed.

Youare the manager. Your assistant — the one in the sidebar, at the small desk beside yours in the office — is your chief of staff: you tell it what you want, and it hands the work to whoever on your team it belongs to. It is not one of your agents, it does not take up one of the desks your plan pays for, and it never raises anybody's permissions by asking. Two things make that practical:

Agents cannot delegate to each other. Only you can hand out work — an agent can't route around its own limits by asking a more-privileged teammate to act for it.

Your dashboard draws the workspace as a room your agents work in. An agent is at its desk because a task is running, and standing by your desk because an approval is really waiting for you — the picture is your workspace's actual state, not an animation.

You choose how that room looks in Settings → Office appearance. Your workspace is one room, and it can wear five different looks — studio, loft, skyline, control and garden. A preview shows the change as you pick, and it saves straight away. How each agent is drawn belongs to the agent rather than to the workspace: you choose its character when you create it, and can change it later from the agent's own card.

Desks are yours to arrange.In the same place you can name each desk, choose which agent sits at it, reorder them, and add or remove desks — up to eight. Until you arrange them, the office seats your agents in order and names each desk after whoever works there. An agent you don't seat keeps working; it simply isn't drawn in the room, and the room says so rather than hiding it. (More desks — a bigger room — are coming.)

The room answers. Hover anyone and a card says who they are, what they are doing right now — the actual task — how long it has been running, and, when they are waiting on you, which tool they want to run and how risky that call is. Click them and the Team conversation opens addressed to that agent, while the agent answers where they stand so you can see the room heard you. Your own desk scrolls to whatever is waiting for you; an empty desk or the hire sign goes to your roster.

A team with nothing to do lives in the room. An agent whose last run stopped more than five minutes ago leaves its desk, and what it does next is a small life: the break starts at the water cooler or the coffee machine, cup in hand; after a few minutes it settles into the lounge — the couch, the armchair, a bean bag; every so often somebody gets up to water one of the plants; after an hour of nothing at all, whoever's turn it is plays the arcade machine; and for five minutes out of every half hour, if at least two agents are free and the room has a meeting table, they hold their own standup there — stools on the near side, chairs on the far one, everyone facing the table, mid-sentence. It is not a screensaver: every one of these is driven by the same facts the activity line shows you, so a busy office never does any of it, and a room full of agents on the couch is telling you nothing has happened in a while. Work ends a break instantly — the agent stands up and walks back. An agent that has never run stays at its desk, because how long it has been idle is not something we know yet. Move the couch or the cooler and the behaviour moves with them; take them out and the room quietly stops doing the things it no longer has the furniture for.

A draft comes off the printer.When a run stops, its agent spends the next few minutes putting the work away — and where it goes depends on what the run produced. A run that left you something to read sends its agent to the printer to collect the page; every other run sends it to the filing cabinet. So an agent standing at the printer means exactly one thing: that agent wrote something and you have not read it yet. Open the drafts from Needs you, or click the printer. A room with no printer files everything at the cabinet, the same way a room with no whiteboard quietly reads nothing.

Edit the handbook and the room reads it. For twenty minutes after you change a handbook entry, free agents walk to the whiteboard and read — because the handbook is the one document every agent on the roster works from, so a change to it is news for all of them at once. This is the only thing the room does that is not on a clock: everything in the paragraph above happens because time passed, and this happens because something changed. A busy agent stays at its desk, an agent that has just finished a run files it away first, and a room with no whiteboard on its wall quietly does nothing — the same rule the cooler and the couch follow. Your agents can edit the handbook too, and you will see the team read what they wrote.

The furniture seats as many people as it fits. A three-tile couch takes one agent and a four-tile couch two: the end tiles are its arms, and an agent is a tile and a half wide. An armchair takes one, in the middle, between its arms. Whoever sits leans back against the backrest, thighs along the cushion and legs hanging over the front of it — the same for a stool at the meeting table and for the chair at their own desk. So a couch with one agent on it has not lost the other two: it is a couch, at this size, with somebody sitting on it.

Seating is a drag. Pick an agent up and carry them to another desk to reseat them; drop them on an occupied desk and the two swap. The room refuses a seat that is not a seat, and nothing is saved until you let go — a carry you abandon changes nothing.

Move the furniture.The paintbrush above the room turns on decorate mode. Every piece you can pick up lights its floor tile in cyan, and the one under your pointer burns brighter — so what you can grab is what you can see. Grab a piece by that tile and drag: the room shows where it would land, green where it fits and red where it does not, and the piece you are carrying leaves its own tile until you drop it. A drop that would block the way between your desk and any agent's station is refused, because the room is a floor plan your team actually walks.

Desks move too, and desks can go. A desk lights its three tiles in decorate mode like everything else: drag it anywhere it fits, and the room judges the drop the way it judges a couch — with one rule of its own, that you must still be able to walk to every desk including the one you just moved. Pick a desk and the panel offers two things. Reset sends it back to its row, which is where a desk you have never moved stands. Remove takes the desk out of the room, and with it the seat — so it is refused while somebody works there. Move them to another desk first. A desk wears the colour of whoever sits at it and has one way to stand, so paint and turn are not offered on it.

You can paint a piece, too.In decorate mode, click a piece to pick it — it rings in white and the strip above the room names it — then choose from ten colours. Clicking it again puts it down, and so does leaving decorate mode. A colour is set per piece and kept with its position, so a bookshelf you made teal stays teal wherever it ends up — and it keeps the paint even if the spot you dropped it on turns out to be taken. What a colour does not touch: paper, glass, screens, lamp glow, a piece's shadow and the plant greens, because repainting those reads as a fault rather than as paint. Your agent can do it in words too — “make the couch green”.

You can add furniture of your own. The +button in decorate mode opens the catalogue: fifteen kinds in three groups — something green, something to sit on, something useful. Pick one and the room aims: the tile under your pointer previews the piece's whole footprint (a couch claims six tiles — three across and two deep, because a body has to fit on it — an armchair four, a bookshelf two) and shows green or red before you commit. Click to place it. The kind stays armed so a corner takes four plants in four clicks; Escape, or clicking the kind again, puts it down. You can add up to 32 pieces — each one is an obstacle your team has to walk around, so the room stays a floor plan rather than a warehouse.

Pick a piece and the room offers what you can do to it. Clicking a piece in decorate mode opens a panel in the corner: its name, the things it can do, and a palette of twelve colours in three families — warm, cool and quiet — with the one it is wearing ringed and a way to take the paint off again. It used to be a line of text in one corner telling you to pick a colour and a row of swatches in another, which is why nobody found it.

Some pieces turn. A quarter at a time, in place: the piece rotates, the room does not. Turn a couch and it stands across three rows instead of along three columns, and the room knows it — a turn is judged by the same gate a drag is, so one that would push a piece through its neighbour or wall your team in is refused with the reason.

A turned piece shows you its other side— its real one. A fridge turned away from the room is a plain white box; a bookshelf turned round shows its wooden back, not its shelves; the arcade cabinet's screen goes with the cabinet instead of following your eye. Every band of detail on a piece knows which way it faces, so it travels to whichever side that becomes and is simply not painted when it goes behind. The marks on top travel too: the mug on a desk swings round with the desk.

Thirty-two pieces turn, and the button says how many ways each one has — 2/4 on a couch, 3/3 on a fridge. The numbers are measured rather than decided: each piece is drawn in all four orientations and compared pixel for pixel, and an orientation that reads the same as one already offered is left out. So a fridge stands three ways (its two blank sides are the same picture), a partition and a rug stand two, and a couch stands four. No press of Turn is ever a no-op.

The rest are not offered one, for a reason you can see: a stool, a fern and a pendant lamp look the same from every side, and a poster or a wall clock hangs flat against a wall and has no other side to show.

And there is an undo. Reset drops everything you did to a piece — the move, the colour, the turn. One the room ships forgets its override entirely, so it follows the shipped layout again and improves with it; one you added keeps its place and loses only what was done to it. Before this, every drag was permanent the moment you let go.

Anything can come out of the room — what you added, and what the room came with. Remove on a piece you placed deletes it. Remove on a piece the room ships takes it out and keeps a note that you did, because otherwise the floor plan would put it straight back; the tray then grows a TOOK OUT row with that piece in it, one click from standing where it started. It opens by itself as the piece goes, so the way back is never something you have to know about. This was refused until now — an office with a filing cabinet you do not want is not a better office. Asking your agent to reset the furniture puts every shipped piece back at its default spot, including the ones you took out, and leaves everything you added exactly where you put it. The whiteboard, the filing cabinet and the arcade are not in the catalogue: each one is a place in the room — the handbook, the audit log, the easter egg — and two of a place is one place too many.

Plants, lamps, stools, couches, bookshelves, sideboards, the meeting table and most standing equipment all move — 74 of the room's 96 pieces. The walls move too.Art, clocks, the neon signs, the string lights and the whiteboards mark the patch of wall they hang on, and slide along that wall in half-tile steps — a hung piece stays on its own wall and cannot land on top of another. What does not move: desks, your own desk and the hire desk, since an agent sits where its desk's station is and a desk that wandered would take its occupant's chair with it. Your arrangement is saved to your workspace, and Settings → Office appearance keeps the room you chose.

Your office can be shared. In the same settings you can turn on a public link at /o/your-slug: a read-only view of the same live room, with no task text, no agent names and no identifiers — visitors see that your team is working, not what it is working on. Rotating the slug revokes every link you handed out.

Desk signs are painted in a five-pixel-tall font, so a long name is shortened to fit its board rather than run into the desk next door: it drops whole trailing words first, so “Support Agent” is signed SUPPORT. Accented letters are folded to their plain form — Mühür reads MUHUR — because the font has no accents and dropping the letter entirely would be worse. The full name is used everywhere else.

The room follows your clock. After dark your office is drawn the way it would actually be — lit from the desk lamps and the screens rather than through the windows — and it comes back to daylight in the morning on its own. There is nothing to switch on.

Sunset is worked out from your browser's time zone, in your browser. Nothing about where you are is sent anywhere or stored, and you are never asked for a location. Two rooms never change: control is an operations deck with no daylight in it to lose, so it looks the same at every hour.

All of this changes the picture only. The same agents do the same work in every room, at every hour, with the same permissions — nothing about how your team behaves depends on what it looks like. If anything, the night room is the honest one: your agents keep working while you sleep, and that is what it shows.

You can also just ask: “switch the office to the night room” works from the agent sidebar (set_office_appearance, with get_office_appearance to see the current choice and every option). Arranging desks works the same way — set_office_sections replaces the desk list, and get_office_sections shows the current desks and the agent ids you can seat. A desk in that list can carry a tile of its own (tx, ty) — “put the support desk in the front-left corner” — and a desk with no tile stands where the room's own layout puts it.