Part 3 of 4 · 2 min read
Exercises
Easy
Easy 1: Run it (30 minutes)
Run python -m pytest -q and python run_all.py in code/mini-design.
Done when: you report the bottleneck and headroom for both workloads.
Easy 2: Your numbers (30 minutes)
Replace the workload with your own agent's volumes and token counts.
Done when: you report cost per run and the binding limit.
Hint
take steps and tokens from traces, not guesses.
Easy 3: Review a draft (30 minutes)
Write a five-minute design answer for a support platform and run review on it.
Done when: you list the gaps and add one sentence for each.
Medium
Medium 1: Mock interview (3 hours)
Do a timed 45-minute mock for the support platform with a friend or a recording.
Done when: you cover requirements, estimates, architecture, two deep dives, failure modes and first build, and note what you skipped.
Hint
say numbers aloud and write your assumptions in the corner of the page.
Medium 2: Cost levers (3 hours)
Add routing, a cascade and context compaction to design.py.
Done when: you show their effect on monthly cost and tokens per minute.
Hint
a cascade's cost is small + verifier + P(escalate) x large (module 5).
Medium 3: Research agent design doc (4 hours)
Write a two-page design for the research agent with trade-offs.
Done when: it has requirements, estimates, architecture, failure modes, evals and a build-first plan.
Hard
Hard 1: Validate the estimate (8 hours)
Compare the estimate with real traces from an agent you built. (I did not run this, so your numbers are the ground truth.)
Done when: you report the error in concurrency, tokens and cost and explain it.
Hint
the biggest errors usually come from steps per run and cache hit rate.
Hard 2: Third design (6 hours)
Design a different agent product (code review, voice, ops) with the same method. (I did not run this.)
Done when: its bottleneck differs from the support platform's and you explain why.
Hard 3: Teach it back (6 hours)
Present the support design to someone unfamiliar with the level and answer their questions. (I did not run this.)
Done when: you record the questions you could not answer and study them.