Home›Framework

The Autonomous AI Framework

📖 3 min readUpdated 2026-04-18

This is a working handbook for building AI that runs on its own. No theory without a use. No fluff. Every page is here because, at some point, I wished somebody had explained this thing to me like a human instead of a spec.

The material is organized in eight sections, stacked in the order you'd naturally learn them. If you're brand new, read the first few Foundations pages in order. They build on each other, and by the time you finish, everything else in the framework will make sense without a glossary in your other hand.

If you already know the basics, skip to the section you care about. The detail pages are self-contained: each one is written so you can drop in cold, read it, and walk away understanding the concept. No prerequisites beyond what's covered in Foundations.

A quick map:

Foundations is the vocabulary and mental model. What autonomous AI actually is, the autonomy spectrum, and a glossary for every term the rest of the site uses.
Claude is the model layer. The brain doing the thinking. This section covers what Claude is, how to prompt it for agent work (different from chatbot work), when to use extended thinking, and the mechanics of tool use.
MCP is how the model reaches outside itself, the standard that lets an AI talk to Gmail, GitHub, your database, and anything else. This is the single most important technology to understand in 2026.
Plugins are a related-but-different way to extend agents, especially Claude Code. They solve a narrower problem than MCP.
Claude Code is the harness, the runtime that holds everything together. This section is the most opinionated on the site, because how you configure a harness changes everything about what your agent can do.
Agent patterns are the proven blueprints. ReAct loops, plan-execute, multi-agent, evaluation, caching. The patterns that most working agents use, and the failure modes that trip people up.
Going autonomous is how you move an agent from "runs when I ask" to "runs on its own." Headless mode, scheduling, self-monitoring, safety.
Build guides are hands-on walkthroughs. If you prefer learning by doing, skip here and come back.
Tools covers specialized agent capabilities like browser automation and desktop control, when the agent needs to click and type like a human would.

The whole framework is free, evergreen, and will keep being updated as the field moves. If you notice something wrong or unclear, tell me; I'll fix it.

If you're brand new, start here.

Foundations

Claude (the model layer)

MCP. Model Context Protocol

Plugins

Claude Code, the harness

Agent patterns

Going autonomous

Build guides

Tools

docs.claude.comAnthropic developer docs home