Skip to content

Learn

The sections below are the recommended way to learn AgenticOS, in order. Read them as a course: each one assumes the ones before it, and none of them assumes you have read the source.

Get started

You need a running stack and one agent that answers you. About twenty minutes.

  • Install

    Docker Compose, or the services by hand. Five minutes to a stack you can open in a browser.

  • Your first agent

    A provider key, an agent, a published version, a run that cost something.

  • Concepts

    Spec, version, exposure, trigger, run. Five nouns. Everything else is built from them, so this is the page to reread when something surprises you.

Tip

Read Concepts even if you are in a hurry. Most confusion about this platform is one of those five nouns being taken for another — a spec for a version, an exposure for a trigger.

Lost in the console?

The console is the map of every area — what each one is for, and which page explains it.

Build the agent

Now make it good. Each page here is one thing you give the agent, and they are independent — take the ones your agent needs.

  • Choosing a model

    Which model this agent should use — open weights or closed, what drives the bill, and how to change your mind later.

  • Models and providers

    27 providers, model profiles, fallbacks, and what a token actually costs.

  • Skills

    Written know-how the agent loads only when it decides it is relevant.

  • Context files

    Standing knowledge written once and bound to many agents — a glossary, a tone guide, an escalation matrix.

  • Knowledge

    Upload, parse, chunk, embed. Collections, and syncing a Drive folder or a bucket into one.

  • MCP connections

    Any MCP server by URL, and 59 in the picker with OAuth already wired.

  • The sandbox

    Files and a shell, isolated, with a lifetime.

  • Triggers

    A run that happens on a schedule or on an event, with nobody typing.

Put it in front of people

An agent nobody can reach is a draft. This is how it leaves the console.

  • Environments

    staging and production as names pinned to versions, so publishing and releasing are two decisions.

  • Surfaces

    Web chat, a hosted page, an embeddable widget, the HTTP API, Slack, Telegram and Mattermost — one runner behind all of them.

  • The deployment itself

    Its identity, its sign-up policy, its notices. The things that are about the installation rather than about an agent.

  • Deploy

    Getting the platform onto a host.

  • Rolling it out

    Who does what, a realistic first ninety days, what it costs, and the questions your security review will ask.

Keep it under control

The part most agent frameworks leave to you. Read it before you give an agent a tool that spends money or writes to something.

  • Permissions

    Three layers: what a role allows, what a grant widens, what a scope lets a tool reach.

  • Governance

    Budgets checked before the request, approvals decided once, alerts, and an audit trail that keeps the value rather than the row.

  • Secrets and the vault

    One mechanism for every credential at rest, and deliberately no second one.

How-to — recipes

Short answers to specific questions, once you know your way around.

Looking for how to extend the platform in Python — a new capability, a new connector, a new endpoint? That is under Resources.

Where to go next

When you know how the platform behaves and want to know exactly what a setting, a command or a spec field does, that is Reference.