Week 14: Exercises & Checklist

Graded exercises

1. Easy: Run notebooks/01-react-agent-from-scratch.ipynb end-to-end with the mock brain (no API key needed). It prints a scenario score at the end (a number from 0 to 10). Record that number in the Week 14 sheet of the Excel tracker.

2. Standard: Add a fifth tool, list_carriers (returns carrier name, region, on-time rate, and fleet size from zoro.data), with a proper JSON schema. Register it in the tool registry, then add one new scenario that requires it and confirm the trace shows the tool being called with the right (no) arguments.

3. Stretch: Set OPENAI_API_KEY (or OPENROUTER_API_KEY), re-run the 10 scenarios with the real model, and compare the score against the mock. Open the saved trace and find one scenario where the real model and the mock disagree; write a markdown cell explaining why (tool choice, argument format, or refusal behavior).

4. Portfolio: Commit the agent, the trace log, and a short README block stating: the 10 scenarios, the final score, and the exact guardrail thresholds (max steps, cost cap, refusal policy). This is the first inspectable agent artifact of the program (tracked in curriculum/projects/README.md).

Hints

1. Easy: Run every cell top to bottom; the mock brain needs no key, and the final cell prints FINAL_SCORE as a fraction between 0 and 1. Record exactly that number.
2. Standard: Mirror an existing tool's three-part shape (function, name/description, and parameters schema) in the TOOLS registry; list_carriers takes no arguments, so its required is []. Add a scenario whose expect names list_carriers.
3. Stretch: Diff the real run's trace against the mock's for the same scenario ID; look for a different tool, a different args value, or a refuse where the mock acted.
4. Portfolio: A reviewer reads the README block cold, so state the 10 scenarios, the score, and the three thresholds (max_steps, cost_cap, refusal policy) as concrete numbers, not adjectives.

Checklist (mirrors manifest.json + Excel tracker)

• [ ] Mon: Study the agent loop and ReAct (reference/knowledge-base/10-agents-multiagent.md).
• [ ] Tue: Implement the tool-calling loop by hand; wire 2 tools (track shipment, convert units).
• [ ] Wed: Add planning and reflection; log every step, the trace is your debugger.
• [ ] Thu: Add guardrails: refusal, max steps, cost cap; test adversarial inputs.
• [ ] Fri: Use case: the agent solves 10 unseen scenarios; publish the trace log.
• [ ] Sat: Take the Week 14 quiz (quiz.md), pass with 8/10; record the score in Notes.
• [ ] Milestone: Update the Excel tracker; commit the agent and traces.