Working Across Workspaces

A conversation with your AI employee is usually one continuous thread. But sometimes a piece of work keeps coming back (“weekly research roundup,” “track this long-running project”), and the employee doesn’t want to tangle that into the main chat. Instead, it opens a workspace: a separate place for that specific area of work. This page explains when a workspace appears, what it looks like, and how it’s different from the main chat.

When does one appear?

A workspace isn’t something you create from a screen: the employee decides to open one on its own. If it judges that a task is multi-step and likely to come back again, it opens a new workspace for it. If a similar workspace already exists, the employee hands the new task to that one instead of creating a duplicate. One-off errands (a single search, a quick calculation) are handled right in place, with no workspace involved.

If you ask the employee to “keep this as its own thing,” it decides whether the request warrants a workspace and opens one accordingly.

What does it look like?

A card in the main chat: whenever the employee hands work to a workspace, all three platforms show a card right at that point in the main chat, with its progress and plan completion (e.g. 3/5). On web it’s marked with 🌿; on the iOS and Android apps it has a branch icon. If it’s a brand-new workspace, the card reads “Working in the new workspace Name”; if the employee added the work to a workspace that already existed, it reads “Continuing in the Name workspace.” Tapping the card opens that workspace.
The workspace tab row (web): on the web, as soon as at least one workspace is open, a row of tabs appears near the top of the chat screen. “Main chat” comes first, followed by up to 4 of your most recently active workspaces, and “More” at the end. A running workspace gets a green dot, and one with unread messages gets a count badge. Tap another tab to switch to it; tap the tab of the workspace you’re already in to open its menu (“Back to main chat”, “Rename”, “Split this request into a new workspace”). “More” opens the full workspace list. On desktop web, tapping a workspace tab from the main chat opens it in a panel beside the main chat, and tapping that tab again brings up the same menu. Team chat doesn’t have this tab row.
Workspaces and Run history buttons: on all three platforms, the row just above the chat composer has a “Workspaces” button and a “Run history” button side by side (on web they’re in the bar above the composer; on the iOS and Android apps they’re on the left of that row). “Workspaces” opens the workspace list, and “Run history” opens the record of tasks. The “Run history” button gets a dot while a task is in progress. The badge on the “Workspaces” button differs a little by platform:
- Web: the badge shows how many workspaces are open (or, if there are no workspaces yet, how many tasks are in progress). It fills with color when there’s something to look at (running, unread, or waiting on you), and the icon turns red when a workspace needs your response.
- iOS and Android apps: the badge only counts workspaces that are running or have unread messages. If there’s nothing to check, you just see the icon with no badge. The icon changes color when a workspace needs your response (orange on iOS).
The workspace list: the top entry takes you back to the main chat. Below it, every workspace is listed. On every platform, each row shows its name, a one-line status summary, and its remaining task count. Web also shows the message count, the AI performance tier it’s using, running SU, and when it was last active; the iOS and Android apps also show the unread message count and one more line with the work gathered in that workspace. Rows for open workspaces have a way to rename them (a pencil icon on web, “Rename” in the menu button at the end of the row on the apps).
Run history: what’s happening right now: inside a single workspace, the employee works through several tasks one after another, and that record is “Run history.” If you’ve noticed a one-line marker in the conversation like ”✓ Task completed · Name · 4/4,” tap its “View task log” link, or tap the “Run history” button above the composer, to see the details. A task that’s still in progress appears with its plan checklist; if the employee leaves one open without wrapping it up, you can clean it up yourself with “Mark complete” or “Close task.” Tapping “Continue with this” makes your next message automatically target that task.
Run history now shows runs that never used a task, too: a row in Run history used to represent a “task”, so a quick run that never got grouped into one (a single search, for instance) simply didn’t show up. Now every run gets its own row, with its time, run condition, and opening line, and runs that share a task are grouped under that task’s name. Runs that aren’t part of any task are set apart with a dashed marker. Date groups are based on when an item was last updated.
Run history: details of past work: a finished task keeps its result and how long it took. On web, select a row marked View run details to see Reason left by the agent, Steps, and Handed off, when available. This includes runs that handled email and ended without a chat message. View in chat is a separate button that takes you to the conversation where the run started, including a team chat. If no conversation took place, this button is not shown.
A resident progress strip: when something with multiple steps is underway, like a new hire’s work setup, a small strip appears just above the chat composer showing how far along it is. Unlike a warning or an info banner, it’s not a colored alert. It’s a quiet, neutral gray strip, and it stays on screen no matter how far up you scroll through the conversation. Tap it to collapse or expand it. If more than one multi-step process happens to be running, only one strip shows at a time.
A “running” indicator: while another workspace keeps working in the background, the web app tells you with the bar above the chat composer and a green dot on that workspace’s tab. On the iOS and Android apps, the “Workspaces” button above the composer does this job: a badge appears while another workspace is running (on iOS the icon also pulses), and it changes color if one needs your response. Tap it to open the list and jump straight there.
Your message might wait its turn: the employee only ever runs one conversation at a time. If another workspace is waiting on an approval or an answer, a message you send where you are won’t be processed until that’s resolved. On web, a small notice appears above the composer when this happens, and tapping “Respond now” jumps you straight to the workspace that’s waiting. On mobile apps, the message you sent just appears dimmed to show it’s waiting.
A response-needed alert: if a workspace you aren’t currently looking at needs an approval or an answer, a small card appears on screen so you can approve, deny, or answer without leaving where you are. Dismissing it just closes the alert; you can still find that workspace in the list later.
A ring fills around the avatar for repeated work: when the employee handles the same kind of work several times over, a small circular gauge appears around the avatar on its most recent message and gradually fills in. It’s a visual sign of “how far along the employee is in learning this repeated job”. Once the ring fills up, the employee may offer to turn that job into something it follows the same way every time from then on. Accept the offer, and the employee sticks to that set workflow going forward; depending on how often the work repeats, this can even lead to setting it up to run automatically as a scheduled run. See Setting Up Scheduled Runs for more on that. Tapping the badge on the ring opens a list of everything it’s been repeating, along with the option to decide on each one.
A save row under a result: a save offer may appear under the result in the main chat once the employee delivers and finishes the work. On web, Save this way sends the save request immediately. If you have unsent text, confirm with Send save request; your text stays in the composer. Answer any pending question or approval first. Sending the request does not mean saving is complete: Inputs and Steps appear once the employee actually saves it. Run again with new material fills a draft, so attach new material and send it. The iOS and Android apps also have a save row, but the immediate-send steps here describe web. For the whole flow, see Save your first task as a workflow and reuse it.
A Workflows panel: repeated workflows the employee has settled on live in one place: the Workflows item in the ⋯ menu at the top right of the screen. It lists a card for each kind of work, showing how many times it’s been done; opening a card unfolds the ordered steps the employee follows for that job. A job with no set way of doing it yet shows as a dashed-border card, and you can press Draft a save request right from there to fill the chat composer with a save request (nothing happens until you send it). Use the pencil button on a card to rename it, and open a card and tap Edit to change its summary, cautions, steps (add, remove, reorder), and branches directly. On web, expand a step and choose Edit this step to change its title and instructions in place, then Save. To request a change through chat, choose Ask in chat and send the edit request filled into the composer. Where a step offers Do it again, you can save whether it should repeat until it meets the condition and its maximum number of attempts, including the first. You cannot edit steps directly in a locked workflow.

Renaming a workspace

If you don’t like the name the employee picked, you can change it.

Find the workspace you want to rename.
- Web: tap the pencil icon on its row in the workspace list, or, while that workspace’s tab is open, tap the tab again and choose “Rename.”
- iOS and Android apps: tap the menu button at the end of its row in the workspace list (⋯ on iOS, ⋮ on Android), or open the workspace, tap “#Workspace name” on the second line at the top of the screen, and choose “Rename.”
Type the new name and tap “Save.” Names can be 1 to 80 characters.

A closed workspace can’t be renamed. The new name shows up right away in the list, the tabs, and the cards.

Splitting a request into a new workspace

Sometimes the employee adds a new request to a workspace that already exists, when it really should have been tracked on its own. As long as that workspace hasn’t started on the request yet, you can split it off into a new workspace.

From the card in the main chat: tap “Split into a new workspace” under the “Continuing in the Name workspace” card. You stay where you are, and the card changes to “Split into a new workspace” with an “Open” link.
From inside that workspace: on web, tap the workspace tab again; on the iOS and Android apps, tap “#Workspace name” at the top of the screen. Then choose “Split this request into a new workspace.” This option only appears when the most recent request sent to that workspace hasn’t started yet. On web, use “Open” in the confirmation message to go to the new workspace; the iOS and Android apps take you there right away.

The split-off request starts fresh in the new workspace and doesn’t run in the original one. Because the same request never runs twice, it only uses SU once. The new workspace is named automatically from the request, and you can rename it as described above.

If the workspace has already started on the request, it can’t be split. The card shows “Already in progress, so it can’t be split.”

How is it different from the main chat?

The main chat is always open and never closes. A workspace closes itself once the employee decides that piece of work is done, or you can close it yourself from the list.
When a workspace closes this way, the conversation it was opened from (usually the main chat) automatically gets a one-line note (something like “The workspace “Name” has finished”) along with its most recent status summary. You don’t have to open the workspace list to see it; you can just carry on in that conversation with something like “okay, write that up.” This note doesn’t ask for a reply, so it won’t restart the conversation on its own.
A closed workspace doesn’t disappear. It moves to a “Closed” section at the bottom of the list. If related work comes up again, the employee reopens that same workspace and picks up where it left off, carrying forward everything it had already learned. A workspace that’s simply gone quiet for a while is cleaned up automatically the same way and lands in that same “Closed” section, and the conversation itself is preserved either way.
Approvals and questions raised inside a workspace are given a shorter wait before they resolve on their own than ones in the main chat. The one exception is a question that requires you to take the screen yourself (signing in, verifying an account, and similar steps). For those, the employee waits until you actually open that workspace. See Handling Approvals for the details.

Do you get notified when it finishes?

The one-line note described above (see the previous section) always lands in the conversation, whether or not you’re looking at the screen. On top of that, if work you handed to a workspace finishes (or runs into something it needs to tell you about) while you’re not looking at that screen, you get a notification. Tapping it skips the list and takes you straight to that workspace. See Notifications for the full rundown of notification types and settings.

How this relates to your automatic run AI usage limit

Autonomous progress inside a workspace draws from the same automatic run AI usage limit (last 24 hours) as any other automatic run. Once the limit is reached, the workspace list shows an “Automatic runs paused” notice and no new scheduled/background run starts (runs in progress finish). Runs start again as earlier usage drops out of the 24-hour window. Messages you send directly are still handled, but child runs started while handling them can be held by the limit.

On web, review and apply the approval cards that configure tables and criteria in the Workflows panel. In the Criteria tab, recent judgments show whether results met the criteria and let you open the related table.

Ask in chat to pin a workflow, then review its approval card. When an organization owner approves a Publish this version and lock it card for a workflow that can be published to an open work portal, the current version is published and locked. Complete any missing setup or requested confirmation first. In a personal space or a space without an open work portal, only the workflow is locked. Check publishing status in Publish. If a single request email sent by a published workflow has a reply deadline, that step waits until the required replies are ready or the deadline is reached before moving on. The work portal indicates that it is waiting for replies and shows the deadline during the wait. This wait does not apply automatically to every email or workflow. Re-running the saved workflow in chat also uses its locked skill versions. If a personal chat points you to a team workflow, open the team task list to run it.

Advanced

This section isn’t needed for everyday use. It’s here for when you want to know how workspaces are created and cleaned up under the hood.

How workspaces are created and revived

The employee has separate tools for opening a new workspace, handing a task to an existing one, and marking a workspace’s work complete. It decides on its own when to use each one; there’s no setting for you to toggle. Completing a workspace also closes any child workspace nested under it.

When the closing note does and doesn’t appear

It only appears for an explicit close: the employee using its completion tool, or you closing a workspace yourself from the list. A workspace that’s simply cleaned up for going quiet (like the 7-day auto-archive) doesn’t get this note. Its conversation is still preserved, just without the note.
When closing a workspace also closes a batch of nested child workspaces underneath it, only the one that was explicitly closed gets a note; its children don’t each get their own. This keeps a single tree-close from flooding the conversation with several lines at once.
The note’s content is pulled directly from that workspace’s name and its latest status summary. If neither was ever set, no note appears at all.
If anything goes wrong while leaving this note, the workspace still closes normally. It’s a courtesy, not a required step.

Nesting and concurrency limits

Nesting: a workspace can open one further workspace beneath it (two levels below the main chat, total). A workspace at that second level cannot open another one; it’s told to route the task through the current workspace instead.
Concurrency: an employee can keep at most 20 workspaces open at once. Once that cap is exceeded, the oldest idle workspaces with no open tasks are automatically archived. A workspace with an open task, an active scheduled run, or an unfinished child workspace is never archived this way.
Repeat-dispatch guard: to stop self check-ins from repeatedly handing work to the same workspace, there’s a minimum 24-hour gap between self check-in dispatches to it. Direct requests from you aren’t subject to this gap.

Dormancy, archiving, and deletion timeline

A workspace with no recent activity is cleaned up in stages:

After 7 days with nothing to do, it’s automatically archived (moved to cold storage).
A workspace that finished or failed is archived the same way after 24 hours.
Archiving is non-destructive: the conversation is preserved, and a related request instantly reactivates it right where it left off.
Once a workspace has been archived for 90 more days, its conversation is permanently deleted and can no longer be recovered.

The unit behind each Run history row

A Run history row used to only appear once a task was created. Now every run gets its own row, including short ones that never used a tool, and runs that belong to a task are grouped under that task’s name. Date groups are based on when each item was last updated. Web, iOS, and Android all show it the same way.

When the conversation jump doesn’t show up

The “View in chat” button in Run history jumps to the message logged when that run started. Where to jump is recorded the moment the run is saved, so it lands correctly even when several things were happening at once. If nothing was ever said inside that run (it was created and closed without any conversation), there’s nowhere to jump to, so the button doesn’t appear at all. A handful of very old records predate this, so for those the button jumps to a guess near when the run ended instead. On mobile apps, if a task lived inside an old workspace conversation (not the main chat), that history may not be loaded on screen yet, so the button can be missing there too. Go back to the list and check again from the latest state.

On web, View run details is separate from the conversation jump, so you can inspect the record even when no conversation took place.

Approval and question wait-time caps

Approvals raised inside a workspace are capped at 10 minutes and questions at 15 minutes, even when a scheduled run or self check-in multiplier would otherwise extend it further. This cap doesn’t apply to the main chat. The one exception is a question where the employee hands screen control over to you, which keeps its original wait time. These numbers may be tuned as the feature stabilizes. For how to respond to an approval card, see Handling Approvals.

Cost

Usage inside a workspace is billed in SU and counts toward the employee’s automatic run AI usage limit like any other automatic run. You can see it broken out by run condition on the employee page’s Logs subpage.

What feeds the progress strip

Right now only two things show up in this strip: a new hire’s work setup progress, and the step progress of an in-progress task/plan. It sits just below the banner area at the top of the screen, above the message list, and whether it’s collapsed is remembered separately per employee, per device (browser).

How renaming and splitting work in detail

Name rules: leading and trailing spaces are removed and runs of spaces are collapsed into one; the result must be 1 to 80 characters. After a rename, what kind of work that workspace collects is re-classified based on the new name.
Which requests can be split: every time the employee hands work to an existing workspace, a request record is kept. The card’s “Split into a new workspace” button only appears for requests among that workspace’s 5 most recent, and the split option in the workspace menu only targets the single most recent request.
Why it never runs twice: splitting and starting the run compete for the same request record, and whichever gets there first wins. If the split wins, the original workspace skips the request; if the run already started, the split is refused. Splitting a request that was already split just returns the same new workspace.
Requests in a closed workspace: closing a workspace doesn’t cancel requests already queued for it, so a not-yet-started request in a closed workspace can still be split from its card.
The new workspace’s name: taken from the start of the request’s first line (up to 40 characters).

How the repeated-work ring counts

The “3” in the badge’s n/3 isn’t a repeat count. It’s the three stages a single piece of work passes through (started → in progress → done). Whether two pieces of work count as “the same kind” isn’t decided by which tools were used or in what order. It’s the task type the employee itself attaches when it marks that work done. Only work from the last 30 days counts, and choosing “Not this” in the ring’s dialog stops that task type from being counted or suggested again. How often the same task type repeats can also feed into a suggestion to set it up as a scheduled run.

What you can and can’t edit from the Workflows panel

The employee writes a workflow first during conversation. After that, you can edit its name, summary, cautions, steps, and branches directly in the panel, or ask for changes in chat. You can also remove items from the inputs or change whether they’re required in the editor. Adding items and settings such as who confirms a step, judging criteria, and locks can only be changed through chat, and a step that carries one of these settings can’t be removed in the editor. Items picked from a table by condition are always optional, and for items that take a table value, the linked table field’s setting decides whether it’s required, so neither can be changed in the editor. A locked workflow can only be renamed. The employee may still refine the workflow as it keeps doing the work. A published portal function picks up the new name the next time it is published. If a saved workflow’s branching structure runs into a problem, the panel flags it and falls back to showing the steps as a plain ordered list until it’s fixed.

Managing Employees on the Home Screen: see recent work and progress for every employee at a glance
Setting Up Scheduled Runs: how this differs from waking an employee at a set time
Handling Approvals: how approvals and questions are handled inside a workspace
Notifications: notification types and settings
SU & Billing: how the automatic run AI usage limit works
Managing Your AI Employees: turning scheduled runs and autonomous activity on/off, checking cost