Agentixit Offers a Local-First, Code-Driven Guide to Building AI Agents

Summary

Agentixit is an open-source, textbook-style teaching repository for practising engineers who want to build production-oriented AI agents. It combines concept-first chapters with small, self-contained code modules that run locally through Ollama, and maps the lessons to examples covering LLM basics, prompting, tool calling, RAG, memory, and agent execution. The guide explains an agent as a goal-oriented system in which an LLM supplies reasoning and planning while tools and retrieval systems perform actions and provide external information. It describes the controller loop as observe, reason, decide, act, reflect, and repeat, then examines prompt chaining, ReAct-style execution, and sequential or parallel multi-agent patterns. For knowledge applications, it compares basic and advanced RAG, including chunking, vector search, hybrid retrieval, query rewriting, reranking, knowledge graphs, and evaluation of recall, relevance, faithfulness, and answer quality. Production chapters discuss LangChain abstractions, LangGraph stateful workflows and checkpointing, and conceptual differences among LlamaIndex, DSPy, and Embabel. They also address short- and long-term memory, context-window limits, latency, cost, observability, hallucinations, bias, evaluation benchmarks, prompt injection, guardrails, sandboxing, and human review. A final section introduces MCP as a standardized way to discover and invoke tools, resources, and prompts through local stdio or Streamable HTTP, and contrasts it with A2A delegation between agents. The text is licensed CC BY-NC-SA 4.0, while the code and docs are MIT-licensed.