1 hour ago
Development
[100% OFF] Agentic AI: From Chatbots to Agents That Get Work Done
Learn how AI agents work, build them in n8n and Python with MCP, then evaluate, guard and cost them. Free local labs.
Course Description
This course contains the use of artificial intelligence.
Everyone is selling "agents" right now. Some of those products really do plan, call tools and finish multi-step work. Many are chatbots or fixed workflows with a new label. This course gives you a clear, vendor-neutral way to tell the difference, and then takes you from understanding agents to building, testing and costing one.
It is a broad introduction for professionals and new builders. You do not need a machine-learning background. If you use ChatGPT, Claude, Gemini or Copilot regularly and want to know what changes when the model starts taking actions, this is the right starting point.
One company, one pilot, start to finish
We follow Harborline Logistics, a fictional freight forwarder with a backlog of shipment-status emails and late customs paperwork. Its operations team runs a 90-day agent pilot called Relay. Every concept in the course is applied to that pilot, so you see the same agent grow from a first loop to a tested, supervised and costed system.
What the course covers
- Section 1, What Agentic AI Really Is: chatbot, copilot, workflow and agent; the agent loop; five levels of autonomy; how to spot agent washing
- Section 2, How Agents Think: model, tools and instructions; ReAct, planning and reflection; five workflow patterns and when not to use an agent
- Section 3, Tools, MCP and Computer Use: tool calling, the Model Context Protocol, browser and computer use, and agent modes in ChatGPT, Claude and Gemini
- Section 4, Memory, Knowledge and Context: context engineering, retrieval (RAG), skills, and grounding an agent in company procedures
- Section 5, Building Agents Without Code: the no-code landscape, writing an agent spec first, and building a shipment-status agent in n8n
- Section 6, Building Agents in Python: choosing an SDK without lock-in, a working agent in about fifty lines, connecting it to an MCP server, and reading traces
- Section 7, Multi-Agent Systems: when one agent is not enough, orchestrators and handoffs, the A2A protocol, and a two-agent customs desk
- Section 8, Trust, Value and Adoption: evaluation, common failure modes and guardrails, human approval, cost per task, KPIs, and a 90-day adoption roadmap
Hands-on labs that run on your own computer
The build labs run locally with free tools: Python, Ollama for local models, and n8n in Docker. No paid API key is needed. The starter folder includes a mock shipment-tracking API, an MCP server, twelve Harborline procedures, and the test emails and customs cases used in the videos.
- Lab 1: run the same task in the agent modes of ChatGPT, Claude and Gemini and read their action logs
- Lab 2: index the Harborline procedures and compare answers with and without retrieval
- Lab 3: build a no-code status agent in n8n and run it against ten test emails
- Lab 4: rebuild that agent in Python, move its tool behind an MCP server, and find a planted bug from the trace
- Lab 5: add a second agent with a handoff and compare results and token use with the single agent
A few honest notes. Lab 1 uses the agent mode of a chat assistant in your browser, and which agent modes you can use depends on your plan and region. If yours does not offer one, you can still follow the recorded sessions. Local models are not deterministic, so your answers and pass rates will differ a little from the videos. The local model used in the labs needs about 6 GB of free memory. The Python section is optional if you only want the concepts and the no-code path.
Practice and assessment
Each of the eight sections has a quiz and a practical assignment with a worked model answer. The capstone asks you to pick one task from your own job and produce an agent spec, an autonomy level with approval gates, a ten-case evaluation set, and a cost estimate with a 90-day roadmap. A 60-question final practice exam covers all eight sections.
What this course is not
It is not a deep engineering course on a single framework, and it does not promise a job, a certification or a particular result. It is a map of the field with enough hands-on work that you can judge agent proposals, run a small pilot, and decide what to learn next. Tools in this field change quickly, so the course teaches portable pieces (spec, prompts, tools and evaluations) that carry over when a platform changes.