All-in-one document to learn AI agent development, stay updated, build real agents and more
A step-by-step guide to becoming an AI agent developer — the frameworks, protocols, memory systems, and infrastructure that separate a real production agent from a demo.
Roadmap (Things to Learn)
- Biblography
- Programming Languages & Foundations
- LLM Providers & APIs
- Agent Frameworks & SDKs
- Tool Use, Function Calling & MCP
- Browser Automation & Computer Use
- Memory & Context Engineering
- Vector Databases & RAG
- Sandboxing & Execution Environments
- Observability, Evals & Compliance
- Deployment & Cost Management
- Version Control
Projects & Preparation
What Is It?
The AI agent developer roadmap on iHateReading connects frameworks, protocols, memory, sandboxes, observability, and deployment into one path.
It targets engineers moving from demo chatbots to production agents that use tools safely and can be audited.
Each page mixes core concepts, videos, and links to deeper blog posts — this hub orients you across chapters.
Why Should I Care?
Job postings want agents that do work, not wrappers around a single completion call.
Production agents fail on memory, tool auth, and cost — chapters map directly to those incident types.
Real World Analogy
This roadmap is a flight plan from 'I can call an API' to 'I ship agents' — each chapter is a leg with fuel (skills) you need before the next hop.
Visual Mental Model
foundations -> frameworks/tools
-> memory/RAG
-> sandbox -> observe -> deployBeginner Explanation
Follow the sidebar order the first time: programming basics, agent frameworks, tools (MCP), memory, vector DBs, sandboxes, observability, deployment. Come back to chapters when your project hits that wall.
Smallest Possible Example
print("Agent loop: think -> act -> observe -> repeat")Common Mistakes
❌ Jumping to LangGraph before Python and HTTP basics.
You cannot debug tool loops if imports, async, and REST still feel magical.
❌ Treating memory and sandboxing as 'later problems'.
The first time your agent runs user-supplied code without isolation is the last time security trusts you.
❌ Skipping observability because demos 'look fine'.
Multi-step failures hide in traces — you will fly blind in prod.
Real Project Example
Intercom, Salesforce Einstein, and Microsoft Copilot extensions all combine models, tools, and policy layers like this roadmap describes.
Startups shipping on OpenAI Agents SDK and Pydantic AI follow the same stack shape.
Build This Mini Project
Draw your current project on paper: model, tools, memory, sandbox, logs. Mark which roadmap chapter fills each blank box.
Interview Question
What layers does a production agent need beyond the LLM?
Tool interfaces (often MCP), memory/RAG, execution isolation, observability/evals, deployment and cost controls, plus governance — the model is one component in a system.
Key Takeaways
- Use this hub as ordered checklist.
- All five production layers matter, not just the framework.
- Revisit chapters when building, not only when reading.
- Python-first examples match most agent SDKs.
YouTube learning
Core concepts from video walkthroughs:
