Deploying Agents

The capstone is a whiteboard, not a repo

You are on holiday with barn wifi. You are not deploying anything. So the capstone is the thing that actually predicts an exam score: draw the whole deployment from memory, then defend every box.

Take a blank page. Draw a control plane and a data plane, and draw the arrow between them pointing the correct way. Inside the data plane, draw one deployment. Inside the deployment, draw two graphs. Hang a default assistant off each graph, then hang a second, differently configured assistant off one of them. Draw a thread. Draw a run connecting an assistant to that thread. Draw Postgres underneath the whole thing and label everything it is holding. Draw Redis next to it and label what it is holding, which is much less than you think.

Then defend it

For every box and every line, answer out loud:

1. What versions this thing, a revision or an assistant version? What does the other one version instead?
2. If I change a secret, which boxes have to be rebuilt?
3. If this deployment is in the EU instance and I want it in the US instance, what do I do? (The honest answer is short and unhappy.)
4. A request arrives. Which handler family runs first, and what are the four things the second family can return?
5. The server restarts. Which boxes on my drawing still contain data, and which are now empty?
6. A user sends a second message while a run is in flight. Name the four strategies, and say which one throws away work already done.

The bar

You can do the drawing in under five minutes, from memory, and you get all six defences right without hedging. Miss one, go back to the module it came from and reread just that lesson. This is the whole domain, compressed into a page you can redraw in the exam room on scratch paper before you answer a single question.