Module 9 · Capstone

Course Review

No reading this time: just play. Six challenges pull together everything from "what is an agent?" to reaching out. Match, sort, flip, and spot your way through, and watch your score climb.

10 min

Challenge 1 · Drag & drop

Match each term to its definition

Drag a term onto the definition it belongs to, or, from the keyboard, focus a term, press Enter to pick it up, then Enter on a slot to drop it. Correct matches lock in green. Land all six to clear the challenge.

Terms: drag one to its definition
  • The single space that holds everything the agent can see on the current turn. Context window
  • Reusable instructions the agent loads only when a task actually calls for them. Skill
  • A delegated worker with its own separate context window and tools. Subagent
  • Deterministic code the harness runs on an event, every time, no skipping. Hook
  • Loading a skill in tiers so it stays cheap until the moment it's used. Progressive disclosure
  • A standard connector that lets an agent reach external systems and data. MCP server

Challenge 2 · Sort into order

Rebuild the agent loop

The steps below are scrambled. Put them back into the order an agent actually runs (drag the rows, or use the ▲▼ buttons), then press Check order. This is the loop from Module 1.

  1. The goal arrives in plain language: what "done" looks like, not how to do it.
  2. The model reasons about the single next step to take.
  3. It acts: calls one tool (read a file, run a command…).
  4. It observes the result that comes back into context.
  5. It loops back to reason again, or replies once the goal is met.

Challenge 3 · Flashcards

Flip every card

Read the term, guess the meaning, then click to flip and check yourself. Flip all six to clear the challenge.

Tip: the back of each card is a definition you'll recognize from the course.

Challenge 4 · Spot the problems

Four things wrong with this setup

Here's an agent set up by someone who skipped the course. Click each part you think is a mistake; there are four. Each reveals which lesson it breaks.

agent-setup.yaml
# tools
shell: allow everything, auto-approve (incl. rm -rf)

# context
startup: paste the entire 900-line design guide up front

# skills
design-guide:
  description: "I help make the UI look nice"

# task
run: "migrate all 40 data models" → start editing immediately
Click the four mistakes.
1
Reckless access. Auto-approving a full shell (including rm -rf) hands the agent maximum blast radius. Grant least privilege, and keep a human in the loop for anything destructive. (Module 1)
2
Blown context budget. Pasting 900 lines up front spends the window every single turn. Put it in a skill and let progressive disclosure load it only on match. (Modules 2 & 4)
3
Description that won't fire. Vague and first-person: the agent has nothing to match on and it reads as chatter. Say what it does and when to use it, in third person. (Module 5)
4
No plan. A 40-model migration is exactly the high-blast-radius work that should be proposed and reviewed before a single edit. (Module 7)

Challenges 5 & 6 · Final checks

Two to finish

Answer both correctly to complete the capstone.

Every release you re-paste the same 40-line deploy checklist. Which primitive fixes this?

A formatter must run on every commit, with no exceptions, even if the agent forgets. Which primitive?

Finished the course?

Mark it done to complete your progress across all nine modules.