1. Courses
  2. Building AI Agents
  3. Capstone

Capstone project for Building AI Agents

Ship an operations assistant

Mangrove's store managers spend part of every morning chasing short deliveries, searching procedures and raising follow-ups by phone. Build an assistant that answers their questions from a set of versioned store procedures with citations, looks up store data through at least two read tools, creates follow-up tasks, and raises at least one risky action (such as a stock transfer request) only with the manager's approval. Use a sandbox with sample stores, stock, deliveries and procedures that you create. Prove with an eval set of at least 30 golden tasks, run several times each, that it is safe to pilot in five stores, and write the pilot plan with budgets, gates and stop rules.

What you hand in

  1. A repository with the working agent: at least two read tools, one write tool behind approval, a policy layer, durable task state and a README
  2. An eval set of at least 30 golden tasks in a sandbox, with outcome and trajectory checks and results per category over repeated runs
  3. A pilot plan for five stores: stages, entry and exit gates, stop rules, budgets per task and per day, and who reviews what
  4. A cost and latency estimate per task from measured traces

Check your work

Tick each item when your project meets it. Be strict: this is the rubric the reviewer uses.

  • Every tool has a written permission rule; the agent acts as the signed-in manager with scopes from the session, never with a shared admin key
  • Risky actions go through a policy layer and an approval card that shows the exact effect; retried steps cannot repeat an action (idempotency keys)
  • Answers from procedures cite the procedure, version and step, use only versions in effect, and say so when the procedures do not cover the question
  • Task state lives in the database, so a restart or a long wait for approval resumes without losing or repeating work
  • Eval results are reported per category with outcome and trajectory checks, over several runs per task, including injected instructions and out-of-scope requests
  • Cost and latency per task are estimated from traces, with budgets and what happens when a budget is hit
  • The pilot plan has stages with numeric gates, stop rules, a kill switch for write tools and an audit trail a store manager can read

0 of 7 checked

Model answer

The model answer opens once your capstone passes the review, so you can compare your work with an outline of a strong submission.

Hand in your capstone

Your links are checked automatically, then the work is scored against the rubric above, item by item, with comments. A pass counts toward the course certificate and shows on your profile.

Loading…