[Oreilly] Building Agentic AI From Workflows to Production [03/26, RUS, ENG]

Lesson 1: Production-Ready AI Workflows
This lesson establishes the foundation: what AI workflows are, how they differ from agents, and why evaluation should be built in from the start. Youll build a complete text-to-SQL workflow with RAG using LangGraph and the BIRD benchmark dataset, indexing evidence into a vector database, retrieving relevant context via embedding similarity, and generating SQL queries from natural language. Youll then evaluate retrieval quality with precision and recall, test SQL correctness against ground truth, and compare six LLMs head-to-head on accuracy, latency, cost, and success rate.
Lesson 2: From Workflows to Agents
Youll convert the text-to-SQL workflow from Lesson 1 into a fully autonomous ReAct agent equipped with three tools: evidence lookup, SQL execution, and schema retrieval. Along the way, youll learn how tool-calling LLMs work, how agents reason in loops, and how real-world agents like ChatGPT are structured. A major focus is evaluation: youll measure agents using explicit metrics (tool selection accuracy, cost per task, token efficiency) and implicit metrics (drop-off rate, copy rate), and apply rubric-based LLM grading.
Lesson 3: Agentic Systems in Practice
When a single agent isnt enough, you need multi-agent architectures. This lesson covers six design patterns: sequential, aggregator, routing, any-to-any, supervisor, and supervisor (tool-calling) with the tradeoffs of each. Youll build a complete multi-agent AI SDR (sales development representative) using MCP and Docker, with agents handling web research, CRM actions, and email outreach. Youll also run agent prompt-engineering experiments on an Airbnb policy compliance dataset, discovering that a single two-sentence prompt addition can dramatically improve tool use and accuracy. The lesson closes with a deep-research hybrid workflow combining planning agents, replanning, and agentic execution.
Lesson 4: Multimodal, Reasoning, and Coding Agents
This lesson moves beyond text. Youll build coding agents where tool calls are raw Python code rather than JSON payloads, using Moondream for visual understanding. Youll benchmark reasoning models and see how reasoning might sometimes improve accuracy on visual tasks but at significant speed and cost tradeoffs. Youll construct a computer-use agent using both multimodal and DOM-grounded approaches, and build a production voice agent with Twilio, Groq Whisper, Llama-4-Scout, and TTS.
Lesson 5: Future Directions in Agentic AI
The final lesson tackles what it takes to run agents in production at scale. Youll implement long-term memory architectures inspired by the Extended Mind Thesis, giving agents a notepad tool and measuring whether self-recorded learnings improve future performance (they do). Youll explore the three layers of guardrails (input, prompt, output), implement constitutional constraints, and examine MCP security risks including tool poisoning and command injection.