Track A
CLI Power User
Master existing CLI Agent tools (Claude Code, Codex, Gemini CLI). Best for PMs, knowledge workers, and anyone who wants quick leverage.
CLAUDE.mdSlash commandsMCPGitHub ActionsCost tracking
2026 · Open learning map · v2026.08.14
Agent Atlas
A Structured Guide to AI Agents
A curated learning map: 9 stages, 23 hands-on exercises, 240+ resources. From Prompt Engineering to Harness Engineering — zero to production-ready agents.
9
Stages
23
Exercises
240+
Resources
77+
MCP / Skills
Pick your path
Two learning tracks
After shared Stages 0–2, choose the CLI Power User track or the Agent Builder track.
8–10 weeks
Track B
Agent Builder
Build your own Agent stack from scratch: tool use, frameworks, RAG, multi-agent systems, and deployment. For developers and researchers.
Function CallingLangGraphCrewAIRAGFastAPIDocker
Build your first agent in 7 steps →By role
Five role-based paths
Different careers need different Agent skills — here are focused paths for five typical roles.
💻
For developers
You can code and want production Agent systems. Focus: tool use, frameworks, deploy, observability.
Track B primary
🔬
For researchers
Academic or industry research focus on eval and frontiers. Focus: eval methods, papers, reproducible experiments.
Track B + 7.5
📊
For knowledge workers
PMs, analysts, operators. Focus: CLI leverage, MCP + Office, daily automation.
Track A primary
🎓
For teachers
Educators bringing AI into class. Focus: Skills design, materials, feedback analysis.
Track A + Stage 5
👥
For everyone
No coding background — learn to use AI tools well. Focus: prompts, CLI basics, pitfalls.
Stages 0–2 + Track A
Core framework
Five layers of concepts
From writing prompts, managing context, autonomous loops, and graph workflows to programmable harnesses.
Stage 7
Harness Engineering
Make agents reliable: eval, observability, deploy, and guardrails. The harness itself is programmable.
Read the Stage 7 five-layer model →Full curriculum
Stages 0 → 8
Expand each stage for exercises, resources, and goals. Open any card to read the full guide.
Learn: Python basics, CLI, Git, APIs, and JSON. Get your environment ready — this is the starting line.
Do: Install Ollama → pull gemma4:e4b (~7.5GB) → run your first local LLM chat.
View full doc →Standard workflow
Seven steps per exercise
Every exercise folder follows the same structure. Work in this order for the best results.
- 1
Read the README
Pick your language edition and understand the goals and expected outcomes.
- 2
Install deps
Read requirements.txt and run pip install -r requirements.txt (openai + anthropic SDKs).
- 3
Try first
Important: starter.py is a full solution, not a TODO skeleton. Attempt it yourself first.
- 4
Compare
Read starter.py (Path A: Ollama), run it, and check the self-verification block.
- 5
Run tests
Run pytest test.py. Tests are mocked — no live API calls required.
- 6
(Optional) Cloud compare
Read starter_anthropic.py (Path B) and compare local vs cloud implementations.
- 7
Go deeper
Follow README callouts to hello-agents, Anthropic Cookbook, and other deep dives.
Knowledge map
Resource quick index
When you need depth on a topic, use the guides under resources/.
📚
Glossary
30+ terms with stage cross-links.
All stages
🧠
Agent paradigms
Five mental models: ReAct, Plan-and-Execute, Reflection, Multi-Agent, Tool-Use.
Stage 3+
📐
Schema cheatsheet
5 golden rules + 5 anti-patterns for function calling schemas.
Stage 3
🔌
MCP / Skills catalog
77 MCP servers / skills in 6 categories with stars and ratings.
Stage 5
🤖
Subagent cookbook
15 ready-to-use subagent dispatch templates.
Stage 5
🔬
Advanced subagents
Description bugs, composition patterns, debugging tips.
Stage 5+
🍳
Cookbooks
Six 30–50 minute guides: Skills, MCP, Office integrations.
Stage 5
🎓
Online courses
10 reputable AI Agent courses with certificates and audience notes.
All stages
⌨️
CLI Agent guide
Seven CLI agents compared + decision matrix.
Track A
🚀
Setup guide
Zero to Hello World in 30–45 minutes.
Stage 0
🗺
First agent in 7 steps
End-to-end Paper Summary Bot walkthrough (~350 lines).
Track B
🏆
Capstone
One graduation project per track with a 4-level rubric.
Final check
✏️
Style guide
Writing conventions for terms, tone, structure, and code samples.
Contributors
📂
Resources index
Full index of the resources/ directory.
All stages
Cost planning
Learning budget
Three budget options from free to full cloud. Hybrid is recommended — local practice + cloud comparison.
| Plan | Details | Budget |
|---|---|---|
| All-local Ollama | Every exercise on local LLMs — $0 API spend | $0 |
| Hybrid ⭐ | Local Path A + cloud Path B with Haiku | $2–5 |
| All Sonnet | Claude Sonnet for every exercise — highest quality | $20–50 |
Model tip: Stages 1–2 → gemma4:e4b; Stage 3+ → qwen2.5:3b; cloud compare → claude-haiku-4-5.