Agentic AI Course
An agentic AI course is a build sequence for multi-step LLM workflows. DeepLearning.AI's Agentic AI course, taught by Andrew Ng, describes agentic AI as software that uses large language models to complete some or all of the steps in a complex task. An agentic workflow plans a multi-step process, runs it iteratively, and improves the output through reflection and tool use. The same course on Coursera is five modules, marked intermediate, for people who already write Python and know the basics of large language models and APIs. Coursera lists about two weeks at ten hours a week. The order below is the module order on those pages.
Workflows, then the four patterns
Module 1 is Introduction to Agentic Workflows. On the DeepLearning.AI page the lessons are: what agentic AI is, degrees of autonomy, benefits, applications, task decomposition, evaluating agentic AI, and agentic design patterns. An optional setup note, a quiz, and a code example, "Try the research agent," follow.
Coursera's note on the module treats autonomy as a spectrum. The applications it names are invoice processing, customer-service agents, and systems that research and write a report. Task decomposition is how you turn one of those jobs into steps you can implement. Evals are the check on whether a changed step is better. The four patterns saved for later modules are reflection, tool use, planning, and multi-agent collaboration.
The goal, tool, observation, and stop loop is already the subject of how to learn AI agents in 2026. Module 1 starts by naming the steps in a workflow and deciding how you will tell that they worked.
Reflection
Module 2 is the reflection design pattern. The lessons are reflection to improve a task's output, "Why not just direct generation?", a chart-generation workflow, how to evaluate whether reflection helped, and external feedback. The labs are chart generation and SQL generation. A quiz and a graded lab follow.
Coursera describes a pass where the model critiques a draft against explicit criteria, or revises after an error or another runtime signal. The cases it names are revising an email, debugging code, and improving a chart, including a critique of the image itself. The SQL lab is there so the accuracy change can be measured.
Tool use
Module 3 lessons are what tools are, creating a tool, tool syntax, code execution, and MCP. The labs turn functions into tools and build an email assistant that uses more than one call. A quiz and a graded lab follow.
Coursera describes the cycle as detecting a function call, running it, and sending the result back to the model. It says the labs use libraries such as AI Suite, and it includes marker-based calling alongside native tool-calling syntax. Code execution is the case where the model writes code and runs it. MCP, the Model Context Protocol, is introduced as a standard for reaching external tools and data.
How to learn Claude Code is the product path on top of that cycle: Anthropic's quickstart on a real project, one small change, then a commit.
Evals you can act on
Module 4 is Practical Tips for Building Agentic AI. The lessons are evaluations, error analysis and what to fix next, further error-analysis examples, component-level evaluations, how to address the problems you find, latency and cost, and a short summary of the development process. The lab adds a component-level eval to the research workflow.
Coursera names invoice processing and customer-email replies as the traces you read. It separates an end-to-end result from a score on one piece, such as web search inside a research agent, and it names precision, recall, and F1 where a component has that kind of answer. The fixes it lists are the prompt, the model, hyperparameters, and which provider you call, weighed against quality, latency, and cost.
Planning, then more than one agent
Module 5 on the Coursera listing is Patterns for Highly Autonomous Agents. The lessons are planning workflows, creating and executing LLM plans, planning with code execution, multi-agent workflows, communication patterns for multi-agent systems, and a conclusion. The graded assignment is an agentic workflow. DeepLearning.AI names planning and multi-agent work on the course page. Coursera prints the lesson titles.
A plan, on that page, is something the model emits and then runs, either as structured JSON or as code. The examples are a retail agent and spreadsheet analytics, and the two formats are a tradeoff among flexibility, security, and predictability. Multi-agent work follows: specialized agents with different roles and tools. The communication patterns named there are a linear pipeline, a hierarchical manager with workers, and all-to-all messaging.
DeepLearning.AI's page says you build each pattern in Python from first principles before you explore frameworks. Neither syllabus page names a library tour. A public course that does walk named libraries as units is the Hugging Face AI Agents course. A posting that asks for Microsoft's associate credential is a different document: Azure AI Apps and Agents Developer Associate, exam AI-103. Ailurn does not administer that exam.
DeepLearning.AI ties a certificate of completion to its Pro membership. Coursera sells a separate certificate experience for the same material. Ailurn does not issue an accredited certificate.
Ask for this sequence
The part that has to stay in order is the work: a reflection pass you can score, a tool whose result re-enters the model, a component eval, then one plan format and one way for agents to pass work.
Ailurn turns a prompt, a PDF, a GitHub repo, or a docs URL into a course you take in the same workspace. It does not watch a YouTube lecture. Open the AI course builder and ask for a course on this sequence: degrees of autonomy and task decomposition, reflection on a SQL or chart draft, one tool call with the result fed back, a component-level eval on a research step, then a JSON or code plan and one multi-agent handoff pattern.