DocsAgents & automations

Agents & automations

Updated Oct 20, 2026

An agent is a configuration - a model, instructions, a set of tools it's allowed to use, and (if you want) a budget and a schedule. Every built-in and installed app is available to an agent as a tool, so an agent can genuinely act on your files, your calendar, your notes, and anything else on your desktop, not just talk about them.

The assistant

The assistant is the easiest way to reach everything else in the system. Type a sentence into the strip and a conversation opens above it: ask a question, ask it to find or summarize something, or ask it to create an agent or an automation for you. It can surface apps, files, agents and automations as cards you can act on directly.

What's included

Hosted models and web search come with every plan, with a monthly allowance and nothing to set up first - open the assistant on a brand-new account and it just answers. Free and Plus draw that allowance from a set of free-tier models; Pro and above get a larger allowance with more model choice. Some included models aren't available when your deployment requires zero data retention; a compatible one is used instead. Add your own key for any provider (Settings > API Keys) any time you want to use that provider directly instead - your key, your billing, no allowance spent.

The model list is live

The model picker (an agent's editor, and Settings > Models) lists every model the AI Gateway currently offers, not a fixed shortlist - so a model that's new, renamed, or repriced shows up without waiting on an update. Each option shows its price per million tokens; the ones marked "Included on your plan" are the ones your hosted allowance actually covers, listed first. Pick anything else and, on the shared allowance, an included model runs in its place - you'll see a note explaining the swap, with a link to add your own key if you'd rather use that exact model directly. Pasting your own key for a provider unlocks every model it offers, included or not.

If the live list can't be reached, the picker falls back to a short offline list so you're never stuck with an empty picker.

Agents

Set up an agent with a model, a system prompt, and the tools it needs. Give it a name and it becomes something you can talk to directly (in Messages) or hand a one-off task to from the strip with @. Every run is recorded with what it cost and what it did, and can be reverted if you don't like the result.

Every agent in one place

The Agents app's Directory lists every agent that exists on your account, not just the ones you built: the assistant, one specialist per app (it already owns that app's tools), anything your team publishes, and - once you run a company - the six departments that run it. Each row shows what it is, where it came from, and whether you've changed it. Message any of them straight from the row, or open one to see what it does.

Make an agent your own

You can't rewrite a built-in, an app specialist, a department, or something your team published - but you can patch one: add to its instructions, add or remove a tool, or point it at a different model, without losing the improvements we ship to the original. Open any entry in the Directory that isn't your own and you'll see its definition, read-only, with your own changes laid on top. Revert any time to go back to exactly what we ship.

Message any agent with `@`

Type @ in the strip and start typing a name - your own agents, the assistant, an app specialist, or a department all show up. @engineering fix the contact page or @notes summarize this week sends that message and opens the run so you can watch it happen, the same way on your phone.

Teach agents a reusable workflow

Skills are small playbooks agents load only when their descriptions match a task. Ask the Assistant to save a repeated workflow, then edit the Markdown file in ~/Skills/<slug>/SKILL.md if needed. Every agent can discover relevant built-in, team, and personal skills. Skills guide how an agent works; they do not add tools or approve actions. See Agent skills for the file format, team governance, and background-run behavior.

Automations

An automation is a rule: something happens (a schedule, an event on your desktop, opening an app, or a manual trigger), and something follows (a notification, opening something, or running an agent). Automations can chain into multi-step pipelines, where a later step uses an earlier step's output.

Tool permissions

Agents act through tools with your approval built in: anything that changes something asks first, unless you've told it not to for that kind of action. Your own provider keys never leave your browser unless you explicitly opt them into the encrypted vault used for runs that continue in the background.