Nagesh Koritala

AI for product managers

AI for product managers.

Short, free lessons on how AI products really work. Not how to train a model: what it does, what your team builds around it, and the questions to ask before it reaches a customer.

The course as one picture. In the middle, the model, taught in lesson 1. Around it, one part per lesson: 2 the loop, 3 tools, 4 subagents, 5 evals, 6 guardrails, 7 observability, 8 cost and speed. Around everything, a dashed frame: the harness, lesson 9. 9The harness: all of it 2 The loop 3 Tools 4 Subagents 5 Evals 6 Guardrails 7 Observability 8 Cost and speed 1 The modelhow it works

The lessons

Nine questions you’ll be able to answer.

Every lesson opens with a moment at a made-up furniture store and asks what you’d do. Write your answer first. By the end of the lesson, you’ll see how close you were.

Part 1

The foundation

The picture every other lesson builds on.

  1. Lesson 1 · 14 minA shopper opens a furniture store’s chat assistant and types: “Will the Harbor sofa fit through a 30-inch door?” Two seconds later, a confident answer appears.Where do you think that answer came from?How an AI product works.What a model does, and the parts your team builds around it.Start lesson 1

Part 2

How agents work

The loop an agent runs, the tools it uses, and when to split the work.

  1. Lesson 2 · 15 minA shopper types into a furniture store’s assistant: “I need a sofa that fits through a 30-inch door and gets here before I move on the 14th.” No single fact on any product page answers that.What does the assistant have to do, in what order, to give her a good answer?Agentic loops.The loop an agent runs, and when it should stop or ask a person.Start lesson 2
  2. Lesson 3 · 16 minA shopper types into a furniture store’s chat assistant: “Is the Harbor sofa in blue in stock? If it can arrive by the 14th, put it in my cart.” The model can only write text.What has to happen for the assistant to handle that, and which part should it never do on its own?Tools and MCP.How an assistant uses tools, and what MCP changes.Start lesson 3
  3. Lesson 4 · 15 minThe furniture store’s team wants its assistant to handle a bigger request: “I’m redoing my living room. Help me pick a sofa that fits and arrives before my party on the 14th.” One engineer proposes a lead agent running three others: a sofa finder, a fit checker, and a delivery planner. Another says one agent with the right tools will do.Which would you choose, and why?Subagents and multi-agent systems.When to split the work across agents, and when not to.Start lesson 4

Part 3

Making it trustworthy

How to know an answer is good, stop what must never happen, and see what really happened.

  1. Lesson 5 · 16 minThe made-up furniture store’s team rewrites its shopping assistant’s instructions to sound friendlier. A teammate tries five questions, likes every answer, and says: “Looks great. Let’s ship it.”What would you want to know before you agree?Evals.Define a good answer, then test for it before you ship.Start lesson 5
  2. Lesson 6 · 15 minThe furniture store’s assistant reads product reviews to answer questions like “Do people find the Harbor comfortable?” One day, a new Harbor review contains a hidden line, written for any AI that reads it, telling it to say the sofa is discontinued and send shoppers to another website.What should stop the assistant from obeying that line?Guardrails.Stop what must never happen, without blocking customers.Start lesson 6
  3. Lesson 7 · 17 minA shopper writes to the furniture store’s support team. The AI assistant told her the Harbor sofa would arrive October 14. It didn’t. The real date was a week later, and she’d taken the day off work. Your engineer checks and says, “No errors in the logs. Everything ran fine.”How would you find out what actually went wrong?Observability.See what your AI feature really did for each customer.Start lesson 7

Part 4

Running it

What each answer costs, how long customers wait, and how every part fits together.

  1. Lesson 8 · 16 minThe made-up furniture store’s shopping assistant has been live for a month. Finance forwards you the first model bill: about three times what your team estimated, though shopper numbers matched the plan.Where do you think the extra cost came from?Cost, speed, and quality.What one answer costs, how long it takes, and the tradeoffs.Start lesson 8
  2. Lesson 9 · 15 minThe made-up furniture store’s assistant passed every test in the lab. Then a shopper writes: “Last week I told you my hallway door is 30 inches. Can the Harbor sofa fit, and can it get here by Saturday?” The assistant replies: “Happy to help! What’s your door width?” It’s the same model that aced the demo.What do you think went wrong?The agent harness.Everything around the model that turns it into a product.Start lesson 9

Take it with you

Nine worksheets.

Every lesson ends with one you can copy:

  1. AI feature brief
  2. Agent loop design
  3. Tool card
  4. Multi-agent design review
  5. Eval plan
  6. Guardrail plan
  7. Weekly AI feature review
  8. Cost, speed, and quality plan
  9. Harness map and launch check

The topic list was prompted by Brij Kishore Pandey’s overview of AI concepts. Every lesson is checked against its original sources, linked at the end of each one.

Ready when you are

Start with
how it works.

One shopper’s question, seven short parts, and the picture every other lesson builds on.

Start lesson 1

Got a question the lessons didn’t answer? Email me. I read every note.

A sentence so far: The Harbor sofa is. The model weighs likely next pieces, shown with made-up bars: 34 inches is most likely, then available, then blue. It picks one, adds it, and does it again. The Harbor sofa is ? Likely next pieces 34 inchesavailableblue picks one,adds it,and again bars are made up, for illustration