Agentic AI Engineering
The systems track: build, evaluate and secure AI agents, from tokens to a production system, in six modules and eight projects.
A six-module course that follows the life of a production AI system: GenAI building blocks (tokens, prompting, tool use, structured outputs, latency and cost), grounding with RAG and context engineering, the agentic leap (architecture patterns, MCP, A2A, LangGraph and the agent frameworks), evals and observability, fine-tuning and local models with LoRA and Ollama, and AI security and production readiness. Eight projects, each with a code track and a no-code track, grow one AI support desk from a triage service into a red-teamed capstone. Updated for November 2026.
Details to know
- Shareable certificate
- Add to your LinkedIn profile
- Assessments
- 25 quizzes · 86 hands-on exercises
- Projects
- 8 projects, each with a choice of tracks
- Taught in
- English
- Recently updated
- November 2026
- Price
- Free, no signup
What you'll learn
- Build LLM features on typed tool calls and structured outputs, and budget their latency and cost before they ship.
- Ground answers in your own documents with hybrid retrieval, reranking and citations, and engineer what goes into the context window.
- Design agents with the least agency that works, expose tools over MCP, delegate across teams with A2A, and orchestrate durable runs with LangGraph.
- Measure quality with golden datasets, calibrated LLM judges, OpenTelemetry traces and CI gates that block regressions.
- Decide when to fine-tune or run models locally, and train a LoRA adapter you can serve with Ollama.
- Defend agents against prompt injection and data leaks with rails that fail closed, red teaming, and governance mapped to the EU AI Act and NIST AI RMF.
Module 01 · GenAI Building Blocks
Tokens, sampling and reasoning, prompting, function calling, structured outputs, the modern AI stack, and the latency, cost and reliability budgets that shape everything. Project 1 builds a triage service.
- How LLMs Actually Work5 hands-onquizconcept18 min
- Prompting That Survives Contact4 hands-onquizconcept18 min
- Function Calling & Structured Outputs2 hands-onquizwalkthrough26 min
- The Modern AI Stack3 hands-onquizconcept18 min
- Latency, Cost & Reliability Budgets5 hands-onquizconcept24 min
- Project 1: Ticket Triage Serviceproject120 min
Module 02 · Grounding: RAG & Context Engineering
Curate the context window, embed and chunk documents, retrieve with hybrid search and reranking, and answer from your own sources with citations. Project 2 builds a grounded help-center assistant.
- Context Engineering4 hands-onquizconcept22 min
- RAG Fundamentals5 hands-onquizconcept24 min
- Build a RAG Pipeline2 hands-onquizwalkthrough35 min
- Advanced RAG Patterns3 hands-onquizconcept24 min
- Project 2: Grounded Help-Center Assistantproject180 min
Module 03 · The Agentic Leap
From workflows to agents: architecture patterns, MCP for tools, A2A for peers, and LangGraph and the agent frameworks for durable orchestration. Project 3 builds an agent behind approval gates; the mid-course project integrates everything so far.
- What Makes a System Agentic3 hands-onvideoquizsandbox20 min
- Agentic Architecture Patterns4 hands-onvideoquizconcept24 min
- MCP — Model Context Protocol3 hands-onvideoquizwalkthrough30 min
- A2A — Agents Talking to Agents3 hands-onquizconcept20 min
- Orchestration: LangGraph and the Agent Frameworks2 hands-onquizwalkthrough32 min
- Project 3: The Support Agent (Tools, MCP, Approval Gates)project240 min
- Mid-Course Project: Multi-Agent Support Deskproject300 min
Module 04 · AI Evals, Observability & Monitoring
Golden datasets, error analysis, LLM-as-a-judge with calibrated pass marks, OpenTelemetry traces, and the production eval loop that catches regressions before users do. Project 5 adds evals, tracing and a CI gate.
- Why Traditional Testing Fails for AI4 hands-onquizconcept24 min
- LLM-as-a-Judge3 hands-onquizwalkthrough30 min
- Observability & Monitoring3 hands-onquizconcept24 min
- Production Eval Loops4 hands-onquizconcept22 min
- Project 5: Evals, Tracing and a CI Gateproject240 min
Module 05 · Fine-tuning & Local Models
When fine-tuning pays off, LoRA and QLoRA in practice, and running open-weight models locally with Ollama. Project 6 moves triage to a local or fine-tuned model if the evals say so.
- Fine-Tuning: When It Pays Off3 hands-onquizconcept22 min
- LoRA & QLoRA4 hands-onquizconcept22 min
- Running Models Locally with Ollama4 hands-onquizwalkthrough30 min
- Project 6: Go Local, Then Fine-Tune a Specialistproject300 min
Module 06 · AI Security & Production Readiness
The threat surface, prompt injection, hallucinations and PII, guardrails that fail closed, red teaming and governance. Project 7 hardens the desk; the capstone ships it.
- The AI Threat Surface4 hands-onquizconcept24 min
- Prompt Injection, Hallucinations & PII2 hands-onquizwalkthrough32 min
- Guardrails & Runtime Checks4 hands-onquizconcept24 min
- Red Teaming, Governance & Production Readiness3 hands-onquizconcept26 min
- Project 7: Red-Team and Harden the Deskproject240 min
- Capstone: Ship a Production-Ready Agentic Systemproject480 min
Do I need to code to take this course?
No, but it helps. Lessons explain each mechanism in plain English, and all eight projects have a Code track (mostly Python) and a No-code / low-code track built mainly in n8n. You finish either track.
Which models and tools does it cover?
As of November 2026: Claude Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5, OpenAI's GPT-6 family, Gemini 4, and open-weight models such as Gemma 4, Qwen3.8 and gpt-oss, with MCP, A2A, LangGraph, Chroma, Ollama, Unsloth, promptfoo and Langfuse. Each lesson teaches the mechanism first, so it stays useful as versions change.
How long does it take?
About 45 hours: roughly 10 hours of lessons and exercises, and 35 hours across the eight projects. One module a week is a comfortable pace.
What will I have built by the end?
A working AI support desk: a triage service, a grounded help-center assistant, an agent that acts through MCP tools behind approval gates, a multi-agent supervisor, an eval suite with a CI gate, a local or fine-tuned model where it pays, and a red-teamed, documented capstone.
Is it free?
Yes, the course is free. The projects can cost a little in API usage; self-hosted n8n and local models keep that close to zero.
How is this different from a prompt engineering course?
Prompting is one lesson here. This course is about the system around the model: retrieval, tools, agents, evals, observability, fine-tuning, local inference and security, and making them hold up in production.