Most of the AI work in a day isn't one long conversation. It's ten small errands that each want to happen whether or not you're at your desk — the morning summary, the inbox triage, the Friday report. A bot is what you hand one of those to. Here's how to set several up, run them at the same time, and keep an eye on all of them from one panel.
What a bot actually is
A bot is a named worker in a workspace with a provider, a model, and a personality — its standing instructions for how to behave. On top of that it gets things a chat doesn't have:
- Files — its own workspace directory, which persists between runs.
- Channels — Slack, Telegram, Discord or iMessage, so it can be talked to where you already are.
- Schedule — cron entries that start it without you.
- Device — an isolated runtime it can execute code in and expose a URL from.
- Webhooks and watchers — so something happening elsewhere can wake it.
- Skills — installed capabilities, from the workspace library.
Every tool it can reach is a permission you granted, and its loop is capped — 25 iterations by default — so a bot that misunderstands a task stops rather than grinding.
Putting a day on a schedule
Open a bot, go to Schedule, and add a task: a cron expression, the instruction in plain English, and where the output should go. That last part is the one people underestimate — a report nobody reads may as well not run:
| Delivery | Where the output lands |
|---|---|
| Activity feed | Show output in Mission Control |
| Messaging channel | Send to a connected channel |
| Send via email connection | |
| Workspace file | Write output to a file |
A schedule you can't try is a schedule you find out about tomorrow, so every task has Run now. Use it once when you write the instruction, read what comes back, then leave it alone. Pause and Resume are there for the weeks you don't want the 7am message.
Scheduled work needs an always-on provider
A bot that runs at 3am can't run on the Claude subscription on your laptop — that needs your machine awake and the companion listening. Point bots at managed credits or an API key, and keep your local Claude plan for chat and code sessions, where you're at the keyboard anyway.
Running several at once
Bots don't queue behind each other. Each one is its own worker with its own instructions, provider, model, files and device, and they run in parallel, on our infrastructure. Which means the useful pattern is not one clever bot — it's a few narrow ones:
- A morning bot that reads yesterday and writes a summary to a file.
- An inbox bot on an email connection, drafting replies for approval.
- A watch bot on a webhook, that only wakes when something changes.
Narrow bots are easier to instruct, easier to trust with permissions, and much easier to debug when one of them starts doing something odd — you know which one, because it only does the one thing.
Within a single run a bot can also fan out on its own: a session may spawn isolated sub-sessions for pieces of work that shouldn't share context, and they hang off the originating session as a tree rather than getting lost. So "several at once" happens at both levels, and you don't have to arrange either.
There is a ceiling, and it's yours to set: the workspace's Gateway screen carries a max concurrent sessions limit — ten by default — along with the iteration cap. Worth knowing it exists before you wonder why the eleventh thing is waiting.
Watching the lot: the Tasks panel
Everything running in a workspace shows up in one place — the Tasks button in the top bar, with the count beside it. It's the one nav entry that isn't a page: it opens over whatever you're looking at, because the point of running something in the background is not having to leave.
The count is honest about everything in flight, not just bots:
- Active bot sessions
- Scheduled task runs currently running
- Code sessions that are busy or waiting on you
- AI builds from the composer
The panel itself is three sections, each there for a different reason:
Response needed is pinned to the top, and everything in it is reachable by scrolling — a question you never see is the worst thing this screen could do. A code session waiting on a reply, a build mid-conversation, work paused for want of credits: all here.
Tasks is the live list. One row per piece of work: a spinner while it runs, the bot's name, its status, which channel it came in on, and — the part worth having open on a second monitor — the last few lines of what it's doing, replaced in place as they land. A tool call shows as the tool it called. You can watch three bots work without opening any of them.
Workspace activity is what happened, newest first, for when you come back to a workspace rather than watching one.
Finished work waits ten minutes for you
Something that lands while you're on another screen is still in the list when you get back. It leaves once you've actually opened it — which is what stops the counter nagging about a thing you've already read.
Open on a bot row goes straight to that session's transcript. The ellipsis beside it is Dismiss — "I've dealt with this" — which takes the row out of the list and off the counter without touching the work itself. Dismissing something that merely ran or finished makes it stay gone. Dismissing a question means "not now", not "not ever": if it asks again, it comes back. The activity screen keeps showing both either way, because dismissing is a decision about your attention, not a rewrite of what happened.
Stopping things
Three levels, smallest first:
| You want to stop | Where |
|---|---|
| One conversation, right now | Pause, resume or end the session on the bot's Chats tab |
| One bot, until further notice | Pause or Stop on the bot itself |
| Every bot in the workspace, immediately | The kill switch, on the Gateway screen |
The kill switch exists because "turn it all off" should never be a matter of finding nine toggles while something is going wrong. Engaging it halts all bot activity in the workspace; disengaging resumes it. Nothing is deleted either way.
If you've not built one yet, the fastest route is to say what you want in plain English on the dashboard — "a bot that emails me a summary of yesterday every weekday at 8am" — and let the setup assistant put the bot, its instructions and its schedule together for you. Then open Tasks tomorrow morning and watch it work.