What is an AI agent for everyday work?
In Atlas, an agent is a job you describe once, for example: every Friday, chase the invoices I have not been paid for. It gets a name, a picture and a schedule. Then it does the legwork across Docs, Sheets, Boards, files, Calendar and Gmail in one run.
It is not a new chat window. It is a small piece of recurring work with rules around it. Every agent gets a face, one of five buddies or your own picture, and agents show up in your contacts, ready to message.
You can also use Agent mode, which is Atlas with everything else put away: one box, one job, and the steps Ask, Run, Done. You can type or speak, and switch between Agent mode and the full app at any time.
What are some everyday examples?
These are examples of jobs you could describe, not agents that come ready-made. Each uses only the parts of Atlas on this page.
- A morning brief. Every morning at 8, a short look at your calendar, tasks and mail.
- An evening review. Every evening at 9, a short look back at the day.
- A week planner. Every Friday, plan next week's tasks into free time around your events.
- Invoice chase. Every Friday, draft a reminder to each person who has not paid, for your approval.
- Mail into tasks. When something arrives that needs you, add a task, marked as coming from the agent.
- Meeting prep. A short brief before meetings with other people, with notes turned into tasks.
- A research digest. Every Friday, look up the topics you name and put a short digest in a Doc.
- A trip planner. A day-by-day plan with a map, using the Trips app.
- A renewals check. Every Friday, look at the Subscriptions app and tell you what renews next week.
Pick one job that you do by hand every week and start there. A small job that runs reliably is worth more than a big one you do not trust yet.
Not every job needs every part. Some agents only read and report. Others write into a Doc, add rows to a Sheet or post an update on a Board card, and you see what changed.
Agents can also use Atlas apps such as Subscriptions, Budget, Goals or Projects, each of which works with your agents.
How do you build an agent?
You talk to it. Describe the job in a sentence, and Atlas fills in most of the setup as you go: the job, when it runs and which apps it uses, so you spend your time on what you want and not on settings. It asks only what it cannot guess. A checklist shows what is ready. Then you press Try it on last week, and the agent runs on last week's work before anything runs for real.
Nothing runs until you turn it on. How agents work has the full walkthrough.
When do agents run?
- Every morning at 8.
- Every evening at 9.
- Every Friday.
- When something arrives.
Agents run on a server, so they keep their schedules when you close the page. Each run keeps a trace of what happened. The Agents screen shows what each agent is doing and when it runs next.
How do you keep agents under control?
Every action has one of four levels, checked by the server before it runs: it does it, it does it and tells you, it asks you first, or it never does it. Paying or buying for you, and deleting anything for good, are never done.
- Permissions page. Every action is listed, with a switch between On its own and Asks me first.
- Approval cards. Email to anyone else and invites with guests ask first, with Approve, Edit and Not now.
- Standing yes. A permission for one named job, with limits on who and how many a day, until you take it back.
- Pause or stop. Pause one agent, or stop them all, from the Agents screen.
- Memory. Short facts Atlas keeps about you sit in one list you can search, edit or forget. Facts picked up from emails or pages wait for your OK before they count.
- Receipts. Activity lists what agents did and what you approved, with Undo next to what can be undone.
- Atlas Shield. Before any text goes to an AI model, Shield swaps cards, bank numbers, SSNs, passwords, keys and codes for stand-ins. It hides the shapes it recognises, so an unusual format can slip through.
A card you do not answer closes after 10 minutes with nothing sent. If a check cannot be made, the step does not run.
No agent can raise its own level or approve for you. The Trust page explains the levels. When you want to try this on your own week, join the beta.