How to Start
Learning AI Agents

Get familiar with what Generative AI is, how it differs from traditional software, and where it is used in the real world.

Understand how large language models are trained, what tokens and parameters are, and how next-token prediction works.

Learn to write effective prompts: system messages, few-shot examples, chain-of-thought and structured output for better results.

Clean, chunk and format raw documents so they're ready for embedding and retrieval — the unglamorous step that makes RAG work.

Combine LLMs with search so they can retrieve accurate, up-to-date information instead of hallucinating from memory.

Store and search embeddings at scale using tools like Pinecone, Chroma and Weaviate — the memory layer behind retrieval.

Interact with LLMs cleanly using frameworks like LangChain and LlamaIndex, or hit the model APIs directly.

Let LLMs call external tools — search, code, calculators or APIs. This is the leap from chatbot to capable assistant.

Learn how agents perceive, plan, reason and act autonomously — the observe, think, act loop that powers everything ahead.

Explore the major frameworks — LangGraph, CrewAI, AutoGen and smolagents — and understand the trade-offs between them.

Create a simple AI agent that performs a real task end-to-end. Build one from scratch, then rebuild it with a framework.

Design how agents think, act and complete tasks. Learn when to use a fixed workflow versus a fully autonomous agent.

Add memory so agents can recall past actions and conversations across sessions instead of starting from zero each time.

Track agent accuracy, performance and reliability. You cannot improve what you do not measure — set up evals early.

Teach agents to think in logical sequences using chain-of-thought and ReAct so they can break hard problems into steps.

Enable several specialized agents to work together — researcher, writer, reviewer — to solve complex tasks as a crew.

Make agents plan ahead, adapt when steps fail, and retry intelligently rather than blindly repeating the same action.

Add filters and validation to keep agents safe, factual and on-policy — before they ever touch a production user.

Connect agents to tools people actually use — Slack, Notion, Gmail — increasingly via the Model Context Protocol (MCP).

Create agents that run on their own — picking up tasks, acting, observing results and updating their plan in a loop.

Equip agents with your own APIs and Python tools so they can do exactly what your use case demands.

Improve speed, cost and error handling — caching, batching, retries and serving — so your agent is cheap and robust.

Host your AI agent so real users can access it — packaging, testing, monitoring and going live.