Discussion: how the workshop ends

Discussion: how the workshop ends

This is the last 20 minutes of the workshop. It’s where everyone shares their journey.

People will arrive here in different places. Everyone has the Foundations. Some of you finished a track. A few did everything. That spread is expected and it does not matter, because the questions below are not about how far you got. They are about what it was like to work with a model.

Stop the keyboards. Close the laptops halfway. Talk.

---

Discussion questions

1. Was working with the model what you expected?

Before today you had a picture of what “building with AI” is like, from demos, from Twitter, from using ChatGPT. Where did the workshop match that picture, and where did they break it? Both directions count: “easier than I thought” (a tool is just a function with a good description) and “harder than I thought” (the model ignored my instruction, invented a price, routed to the wrong tool). The gap between the picture and the practice is the most useful thing to name in the room.

2. What surprised you?

Good or bad, anywhere in your run. The f1 price that was invented and looked completely plausible. The f5 guardrail that blocked a question you thought was fine, or let one through. A tool that fired when you did not expect it, or refused to. Your neighbour’s f6 routing differently from yours on the same prompt. If you got into a track: three reviewers disagreeing in p3, or a retrieval score that ranked the wrong city in r1.

3. What did you learn that you will actually use?

“Which challenge made you say: I could use that at work,” and for what, specifically? “Tools at work” is too vague. “Typed output for parsing our intake forms” is the answer we want. This separates people who learned a framework from people who learned a tool they will reach for next week.

4. Where would you let the model decide, and where would you keep the steps?

You watched the model pick the right tool from a description, with no routing code. If you did the Patterns track you saw both halves: p1 to p4 where you hold the control flow, p5 to p7 where the model does. What is a problem at work where you would let the model decide, and one where you would write the steps yourself? Listen for “data pipelines I’d write by hand,” “support routing I’d let the model do,” “anything auditable I’d hand-write.” Those are the right instincts.

5. What do you trust now, and on what evidence?

f1’s invented price looked right and nothing caught it. f7 swapped the model for a fake you control and proved the guardrail’s branches. What do you check by eye today that a deterministic test could pin down? And where would you want an eval instead, judging the quality of the model’s answer, the one thing a fake deliberately cannot tell you?

6. What will you do next?

The tracks you did not do are take-home, and so are the self-serve tracks (MCP, Resilience, and the rest). More useful than “which track”: what is the first thing you would try to build at work, and which piece of today does it start from?

---

Share-back (the last 5 minutes)

Bring the room back:

1. “Hands up: who is planning to ship something with tools next month?” A quick temperature check.
2. “What is the most surprising thing anyone at your table said?” Take two or three voices. Second-hand surprises travel better than first-hand ones.
3. Now name the arc. “Everyone here built the augmented LLM: an agent, typed output, tools, a guardrail, a test. Whatever you did after that, orchestration, retrieval, a UI, was those same primitives composed. That is the whole field right now: the primitives are small, the composition is the work.”

People leave with their own framing first and the shared vocabulary second.