Agentic Systems Learning Roadmap

Section 1: Understanding Agents (Projects 1–5)

This series takes you from raw model + tools mechanics to a full local‑first RAG agent.

1. Project 1: What Is an Agent? (No Framework)
   
   
   
   • What an agent is: model + tools + loop
   • ChatOllama as the core “brain”
   • Defining tools with @tool
   • Binding tools to the model
   • Detecting and executing tool calls manually
   • Reason → act → observe → respond loop


2. Project 2: Agent Memory (Stateful Agents From Scratch)
   
   
   
   • What “state” means in agentic systems
   • Why stateless agents fail at multi‑turn reasoning
   • AgentState TypedDict for messages
   • Conversation history with Human, AI, and Tool messages
   • chat_node: model invocation and tool detection
   • tool_node: tool execution and follow‑up messages
   • Multi‑turn loops with persistent memory


3. Project 3: Transitioning to LangGraph (Framework‑Based Agents)
   
   
   
   • Why frameworks exist for agents
   • Nodes, edges, and conditional routing
   • MessagesState as built‑in memory
   • Defining a StateGraph
   • chat_node as a LangGraph node
   • Using ToolNode for automatic tool execution
   • tools_condition for routing between chat and tools
   • Running the graph with agent.invoke()


4. Project 4: Types of Agents
   
   
   
   • ReAct agents (reason + act loops)
   • Tool‑calling / function‑calling agents
   • Router agents (conditional branching)
   • Planner–executor agents
   • Multi‑agent systems and collaboration
   • When to choose each agent type


5. Project 5: Local‑First RAG Agent (Chunking, Embedding, Retrieval)
   
   
   
   • What RAG is and why agents need retrieval
   • Local‑first stack: Ollama + FAISS + LangChain + LangGraph
   • Chunking PDFs for retrieval
   • Embedding with a local embedding model
   • Building and querying a FAISS vector store
   • Creating a retrieval tool for the agent
   • Integrating retrieval into the agent loop
   • Streaming responses and iterative refinement

These 5 projects give you a complete mental model of agentic systems—from raw loops to full RAG workflows.