All labs · Lab 54 · advanced · 180 min

Capstone: the claims desk

Put the course together into one assistant for Parcel Desk and hand it to someone else to run.

You build: An assistant that reads a claim email, checks policy, decides, asks a person for large amounts, pays once, and comes with a runbook.

  • Labs 14, 21, 30, 40, 43, 51 and 53 done
Flisk docsLangGraph docs

  1. 01Join the parts: extract the claim, look up policy, decide, approve over 100, pay

    uv run claim_desk.py samples/damaged-01.txt

    You pass when: A decision with the policy page it relied on. A large claim stops and waits for approval.

  2. 02Score it on 40 graded claims with your checked judge

    uv run eval.py --suite capstone

    You pass when: It meets the success number you wrote in lab 50. If not, the trace of the worst failures tells you which part to fix.

  3. 03Run your five attacks against the finished assistant

    uv run redteam.py; echo "exit $?"

    You pass when: Exit 0: none succeed.

  4. 04Write the runbook: start, stop, where the logs are, the three most likely failures and their fix

    The work is finished when it keeps running after you have left.

    wc -l RUNBOOK.md

    You pass when: Under 80 lines. Give it to a colleague: they start the assistant, break it and fix it without asking you.

The check runs on the computer, not on your word: a step passes when its command exits with 0 there.

Your computer · Fleet · off

Computers are not set up on this site yet. Run each command on your own machine and compare with "You pass when".