Practice: Thinking like an architect
Self-check
Section titled “Self-check”Six short questions. Answer each in your head before opening the collapsible. Active retrieval is where the learning sticks.
1. Every agentic system reduces to four parts. Name them and the job of each.
Show answer
A model (reads, reasons, writes), tools (functions the model can call to act on the world), a loop (act, observe the result, decide the next step, repeat), and a context window (the finite space holding everything the model currently knows about the job). Everything else about a system is how these four are arranged.
2. What is the difference between a workflow and an agent?
Show answer
A workflow follows a route drawn in advance: models and tools are orchestrated through predefined code paths. An agent chooses its own route as it goes, directing its own process and tool use. The workflow is repeatable and cheap; the agent is flexible and harder to predict.
3. When does a task genuinely call for an agent rather than a workflow?
Show answer
When the path through the work cannot be known in advance, that is, when the number and order of steps depend on what the work uncovers along the way. If you can draw the steps on a whiteboard and they hold for almost every case, a workflow with the model working inside each step is the better fit.
4. A rule must hold every single time. Where does it live, and why not in the prompt?
Show answer
In the structure around the model: a check the flow cannot skip. An instruction in the prompt is a request the model will follow almost every time, and “almost” is exactly what a must-hold rule cannot tolerate. Prompts are for behavior that benefits from judgment and tolerates variation.
5. Name the three recurring trade-offs this track is organized around.
Show answer
What the model decides versus what code guarantees (decide versus enforce). One agent versus a team (one versus many). What stays in the context window versus what gets fetched when needed (carry versus fetch).
6. Why do dazzling demos so often die in production while plain ones survive?
Show answer
A demo exercises the model; a product exercises the architecture. The demo never meets the edge cases, the must-hold rules, or the context limits that the arrangement around the model exists to handle. Whether the system survives depends on decisions the demo never tested.
Try it yourself
Section titled “Try it yourself”Design exercise: where should judgment live?
Pick one real task from your own work that you might hand to an AI system. On paper or in a note, answer four questions in order:
- Is the path through this work predictable? Draw the steps. Do they hold for nine cases out of ten?
- List every rule that must hold every single time. For each, write “structure” next to it. List the behaviors where variation is fine. Write “model” next to those.
- Would you give this to one agent or several? If several, name the seam: what exactly passes from one to the other, and what could get lost there?
- What must the system carry at all times, and what can it fetch when needed?
You have just produced an architecture sketch without writing any code. Keep it. In the capstone lesson you will build one of these for real, and comparing that design against this first sketch is the clearest measure of what the track taught you.
Flashcards
Section titled “Flashcards”Q. What are the four parts of every agentic system?
A model, tools, a loop, and a context window. Everything else is arrangement.
Q. What is the one-sentence definition of architecture this track uses?
Architecture is deciding where judgment lives: what the model decides, what the code decides for it, and what waits for a human.
Q. Workflow versus agent: what is the dividing line?
A workflow follows predefined code paths. An agent directs its own process and tool use, choosing its route as it goes.
Q. When is an agent the right choice over a workflow?
Only when the path cannot be known in advance, when the steps depend on what the work uncovers. Otherwise the simpler, predictable workflow wins.
Q. What is the difference between an instruction and a guarantee?
An instruction in a prompt is a request the model follows almost every time. A guarantee is a check built into the structure the flow cannot skip. Must-hold rules get guarantees.
Q. What does a team of agents cost you?
Seams. Every boundary between agents is a place where context gets lost, work gets duplicated, and failures hide. Split only when the work itself splits.
Q. Why is more context not more knowledge?
The context window is a budget and everything on it competes for the model’s attention. Piling everything on makes the model worse at finding what matters. Curation beats volume.
Q. What does a demo exercise, and what does a product exercise?
A demo exercises the model. A product exercises the architecture: the rules, seams, and limits the demo never met.
Q. What is the most senior sentence you can learn to say in this field?
“This does not need an agent.” Simplicity first; escalate to an agent only when the task demands run-time flexibility.
Q. Name the three recurring trade-offs of agentic architecture.
Decide versus enforce. One versus many. Carry versus fetch.