Exercises — Module 08: Architecture patterns

8.1 — Reverse-engineer an architecture

Pick a real agent product. Write a one-page description of its architecture, inferred from its behavior and any public docs. Identify the shape (Module 08 §2), the product surface, and at least two trade-offs you think the team made.

Deliverable: one-page architecture doc.

8.2 — Greenfield design

You are the harness architect for a new agent that operates on a corporate inbox (reads new email, drafts replies, flags urgent items). Produce:

  • A chosen architectural shape, with rationale.
  • The agent graph.
  • A list of the first five tools.
  • The first three safety rails.
  • The first three evals you would build.

Deliverable: a design doc.

8.3 — Compare shapes

Pick a task. Implement it two ways: single-agent loop, and router+specialists. Compare on a 20-case eval: cost, latency, quality, code complexity.

Deliverable: the two implementations and a comparison writeup.

8.4 — Observability from zero

Take a harness that has only print-statements. Add:

  • OpenTelemetry tracing with spans for model calls and tools.
  • Structured logs per turn.
  • A metric for task success, cost, latency, per-tool error rate.
  • A local dashboard (anything — Grafana, a spreadsheet).

Deliverable: code + screenshot.

8.5 — Migration drill

Your provider deprecates your model and you need to migrate. Without having access to a real deprecation, simulate one: switch models, see what breaks, fix, document. Now write the playbook you wish you had had.

Deliverable: the playbook.


This site uses Just the Docs, a documentation theme for Jekyll.