What this is
This is the agent-shaped cut of the AI Path: every article about building, extending, scaling, and operating agents, sorted into the order the ideas actually depend on each other. The broader AI stack map covers the whole library - models, search, coding tools, image tools; this page goes deep on the one layer where teams now spend most of their engineering time.
The spine of the map is a single idea: an agent is a model in a loop, and everything about that loop - what it can do, what it costs, how it fails, when to split it into several - follows from the one block of text it sends the model every turn. Read the layers in order and each article has what it needs from the one before.
The agents are the easy part. The window, the boundaries, and the budgets are the system.- the theme this map keeps returning to
The idea
What "agentic" means, and the stateless engine underneath it. Two articles, one mental model.
The concept piece names the parts - goal, tools, loop, autonomy - and separates an agent from a chatbot. If models themselves are new to you, read the LLM article first: an agent inherits every property of the engine it wraps, statelessness above all.
- → What is agentic AI? LEVEL 101
- → What is an LLM? LEVEL 101 - the engine, read first if models are new
The window
One article, and the most load-bearing one on this map. Every layer after it is secretly about the same thing.
Every turn, an agent sends the model one block of text and gets one back. What goes into that block, who puts it there, what each part costs, and what gets thrown away when it stops fitting - this is the mechanical foundation for everything below: extensions manage the window, architectures split it, production budgets pay for it.
- → What's in an agent's context? LEVEL 101
Extending an agent
Functions, MCP servers, Skills, hooks, plugins - five words for attaching capability to an agent, each with a different relationship to the window.
Start with the three-way comparison to get the vocabularies straight, then the protocol that carries tools, then the packaging formats. The four-mechanism piece is the one to send a teammate who keeps mixing them up; the plugins article covers how Claude Code distributes all of it as one install.
- → Functions, MCP, Skills LEVEL 101 - the vocabularies, untangled
- → What is Model Context Protocol? LEVEL 101
- → What are Agent Skills? LEVEL 201
- → Skills, MCP, Hooks, and Plugins LEVEL 201 - the four mechanisms, side by side
- → What are Claude Code plugins? LEVEL 201
Architectures
The step past one agent: whether to take it, and the coordination patterns waiting on the other side.
Read them in this order. The systems piece is the decision: why one agent stops scaling, what handoffs cost, and when a single agent is still right. The patterns piece is the catalog for after the decision: orchestrator-workers, pipelines, fan-out, critique, and routing, with a rule for choosing.
- → Multi-agent systems LEVEL 201
- → Multi-agent orchestration patterns LEVEL 201
Running in production
Building an agent is a week. Operating one is the job. Three articles cover the operating side, and they are best read before the first deployment.
Evaluation is how you know the agent works before trusting it; the failure piece names the recurring ways long runs go wrong; the cost piece explains why a run bills many times what its prompt suggests once the whole context is re-sent every turn.
- → How to evaluate an agent LEVEL 301
- → Why agents fail LEVEL 301
- → What an agent actually costs LEVEL 201
Frameworks and products
The harnesses you would build with, and two shipped agents worth studying as reference implementations - one for engineering work, one for everything else.
The survey covers the libraries and platforms a builder is likely to pick; the Microsoft Agent Framework piece looks closely at one mainstream option. Claude Code and Claude Cowork are the study-worthy products: the same agentic architecture pointed at two different jobs.
- → AI agent frameworks LEVEL 101
- → What is the Microsoft Agent Framework? LEVEL 201
- → What is Claude Code? LEVEL 201
- → What is Claude Cowork? LEVEL 101
Three reading orders
The map is sorted by layer, but most readers want a route. Three orders, depending on where you stand.
From zero to a working loop
The multi-agent route
Costs, failures, and what to buy
- What is agentic AI?
- What an agent actually costs
- Why agents fail
- AI agent frameworks + the Microsoft option
- Claude Code + Claude Cowork as shipped references
The map at a glance
If you read only this section, here is the agentic stack from the idea up:
- Idea: an agent is a model in a loop - and the model is stateless.
- Window: everything the agent knows is one block of text, re-sent every turn.
- Extensions: MCP puts tools in the window, Skills keep instructions out until needed - the four mechanisms in one read.
- Architectures: when to split, then the five coordination patterns.
- Production: evaluate, expect the failure modes, budget for the re-sent window.
- Frameworks and products: the harnesses, and Claude Code / Claude Cowork as the shipped references.
Every article in the map
Every article on the agentic slice, by layer, with a one-line reason to read.
| Layer | Article | Why read it |
|---|---|---|
| Idea | What is agentic AI? 101 | The concept - what makes a system "agentic," the parts of an agent, and how it differs from a chatbot. |
| Idea | What is an LLM? 101 | The stateless engine every agent wraps - read first if models themselves are new. |
| Window | What's in an agent's context? 101 | The one block of text an agent sends every turn - what is in it, what it costs, what falls out first. |
| Extensions | Functions, MCP, Skills 101 | The three vocabularies for "things an agent can do," compared and untangled. |
| Extensions | What is Model Context Protocol? 101 | The open standard for putting tools and data in front of any MCP-speaking agent. |
| Extensions | What are Agent Skills? 201 | Anthropic's packaging format for instructions an agent loads only when the job calls for them. |
| Extensions | Skills, MCP, Hooks, and Plugins 201 | The four extension mechanisms side by side, and which job each one belongs to. |
| Extensions | What are Claude Code plugins? 201 | Claude Code's packaging for sharing skills, commands, and MCP servers as one install. |
| Architecture | Multi-agent systems 201 | When to split one agent into several, the three shapes that work, and what handoffs actually cost. |
| Architecture | Multi-agent orchestration patterns 201 | Orchestrator-workers, pipelines, fan-out, critique, routing - and how to pick the one your job needs. |
| Production | How to evaluate an agent 301 | How to measure an agent before trusting it with real work. |
| Production | Why agents fail 301 | The recurring failure modes of long agent runs, and the guards that catch them early. |
| Production | What an agent actually costs 201 | Why a run costs many times what its prompt suggests once the whole context is re-sent every turn. |
| Frameworks | AI agent frameworks 101 | A survey of the libraries and platforms used to build agents in production. |
| Frameworks | What is the Microsoft Agent Framework? 201 | Closer look at one mainstream option, with the surrounding Microsoft ecosystem context. |
| Products | What is Claude Code? 201 | The agentic coding assistant as a reference implementation - sessions, tools, and the loop in production. |
| Products | What is Claude Cowork? 101 | The same agentic architecture pointed at knowledge work - no terminal required. |