Agents with a job to do.
An agent is an AI worker you set up once: what it is for, what it may reach, and when it must ask a person first. Talk to it in chat, give it duties to do on its own or at the press of a button, approve what it proposes, and see everything it did.
What makes an agent
- A mandate: instructions in plain words (up to 8,000 characters) and the tools it may use, picked as permission sets.
- Access: the workspaces and projects it can reach, and what it may read, edit or delete there.
- Approvals: when it must ask a person before it acts.
- Sharing: who may use it and who may manage it.
- Duties: work it does on its own. See Duties.
An agent never reaches further than the access you give it, and every change it makes is recorded under its own name.
See Security and access for how an agent works for the person who asked, and who reviews its approvals.
Create an agent
Press New agent on the Agents page, which is in the sidebar. The set-up opens over the page in four short pages, and your new agent asks each question itself.
Look and name
Pick one of sixty characters, or an icon on a colour, and give it a name and a line about what it is for. Then write its instructions in plain words (up to 8,000 characters). To start quickly, choose a role: Project Manager, Documenter, Tester or Compliance Officer fills in what it is for and its instructions, and keeps the name you typed.What it can do
Start from a preset, Read only, Editor or Full access, or build your own. What it works on is a row of tiles, one for each kind of thing, and each tile has Read, Edit and Delete. Then choose When it asks you first and Where it can work. See what it works on and where it can work, below.Who can use it
Just you, the people and teams you choose (each to use it or to manage it), or everyone in your organisation. Until you share it, only you and your organisation's owners and admins can see it.Ready
Read what it will be able to do, then create it. Nothing is created before you press Create. Then give it its first duty straight away, or finish and open its page.
What it works on
The tiles are Documents, Diagrams and pictures, Whiteboards, Plans and tasks, Improvements, Databases and dashboards, and the Brain, where your organisation has it. Press a tile to put it in at Read, or choose Read, Edit or Delete. Choose as many as you like.
- Each tile brings every tool that kind of work needs, such as finding the right workspace, folder or person, and search. Read only reads, Edit never deletes, and Delete adds deleting.
- A preset is every tile at one level: Read only is all at Read, Editor all at Edit, and Full access every tool.
- A tile also fills in Where it can work for its kind of thing, unless you set that yourself.
- Choose tools one by one is folded away under the tiles. Open it to give the agent every tool (new ones included) or only the ones you choose, or to add or take off a single tool. Your own changes stay when you change the tiles, and Undo my changes lets them go.
- An agent never gets tools for billing, members, teams, permissions, the Brain's settings, the organisation and its settings, or other agents, whatever you choose.
Where it can work
Choose Everywhere, or Only some places: the workspaces, projects or folders it reaches, each list narrowing the one before. A folder includes every folder inside it. Only owners and admins can give an agent the whole organisation.
With places chosen, the levels can be the same in every place, or different in each place, as for an API key. Databases and dashboards are a tile of their own and never follow Documents: with that tile left out the agent reaches no database, and access limited to folders reaches none either, because a database is never kept in a folder. See Permissions.
When it asks
| Approvals | What the agent does |
|---|---|
| Ask before any change | It reads and answers freely, and proposes every change for a person to approve first. This is where every agent starts. |
| Ask before deleting | It makes ordinary edits on its own, and asks before it deletes or irreversibly changes anything. |
| Don't ask | It acts on its own inside its access. Nobody approves what it does, and every action is still recorded. |
Every level is on every plan. A proposal waits up to 72 hours for a decision. Open Approvals on the Agents page to see what is waiting across all your agents, approve a whole run at once, or decline with a reason that stays on the record. On a keyboard, J and K move between proposals, A approves and D declines.
An agent's page
Each agent has tabs for its Overview, Chat, Duties, Activity and Approvals, then its Mandate, Access and Sharing. The overview shows anything waiting for a person, its approvals and speed, its mandate and access, who owns it, the AI it runs on, and its recent runs. Its Duties tab shows each duty as a card that says, in pictures, how it starts, what it uses and what it makes. See Duties.
Your agents in the top bar
The Agents button in the top bar, beside the bell, opens your agents at a glance: what is working now and the last few runs that finished, the duties you can run now, each with a Run button, and every agent you can use. Press a run to open what it worked on, its approval when it is waiting for you, or its agent. A small number on the button says how many runs are working.
An agent's card on the Agents page has the same Run button when it has a duty you may start by hand; with several, Run lists them. Where a duty has a button on documents, plans, tasks or other items, it is in that item's Agents menu. See buttons on items.
Chat with it
Message the agent from its Chat tab. Turn on Plan first to see its plan before it starts, and set the approvals for this conversation if you want it stricter than usual. You can also talk to an agent from the top bar, by choosing it in the Ask about menu, or from your own AI tools. A conversation carries on while you work elsewhere, and comes back if you reload the page.
Activity and undo
Activity lists every run, how it started (asked, delegated, from chat, from a duty) and what it did. To reverse a run, choose Undo this job: the agent builds an undo plan, nothing changes yet, and the plan is approved like any other.
Pause, archive and delete
- Pause stops an agent at once, until you resume it.
- Archive puts it away without deleting it. An archived agent does not count towards your plan, and you can restore it from the Archived filter.
- Delete permanently only works on an archived agent, and asks you to type its name.
Owners and admins can also adopt an agent that is waiting for a new owner, and download every agent chat and run in the organisation.
Quick responses or thinking
The Speed card on an agent's overview sets how it works on Stable Baseline's AI:
| Speed | What it does |
|---|---|
| Quick responses | Answers fast. Best for everyday questions and simple tasks. Every new agent starts here. |
| Thinking | Takes longer and works through harder problems step by step. |
A change applies from the agent's next run, and only the people who manage the agent see the card. If your organisation runs everyday or thinking work on its own AI, owners and admins choose each agent's AI under Organisation settings, AI instead. On the Free plan, agents work once your organisation connects its own AI.
What agents cost
Asking an agent costs no credits, within a fair-use number of questions a day for each person, shared with Orbit. Work an agent does for a duty uses credits, whether the duty started on its own or someone pressed its button: 1 when it checks and finds nothing to do, 5 when it acts. See what a duty costs. A tool with its own fee costs the same whether a person or an agent uses it.
Limits