An AI agent is a spaceship: a reasoning core, robotic arms (tools), a cargo hold (context), fuel (token budget), shields (guardrails), and a hotline to mission control (escalation). Master each part in a mini-game, assemble them into a working ship, then fly missions where your choices — the same choices the exam tests — decide whether you dock or drift.
Six parts, ~1 minute each. Every part is one exam concept wearing a flight suit. Finish all six to unlock assembly.
Three engines: Haiku (small, cheap, fast), Sonnet (balanced), Opus (deepest, priciest). Fit the cheapest core that can fly each job:
Every transmission to Claude needs exactly 4 pieces. Select the real ones — watch out for parts that don't exist (the API is stateless!) — then transmit:
Capacity 100 units. Load all four subsystems — but the reply needs ≥ 20 units free to come back aboard. Click crates to load/unload; one crate can be split (retrieve only what's relevant):
Two arm blueprints. The metal is identical — only the label the core reads differs. Install one, then run the test request: “where is order #4471?”
"order": "gets order stuff""get_order_status": "look up ONE order
by numeric ID. Use for status/shipping questions. NOT for refunds. Returns status, eta, cancellable."Each engine turn burns fuel. The test task needs ~9 turns. Pick a burn limit and run the test — the core always thinks it's almost done, so the limit is your only real brake:
Policy: refunds over $200 need a human. An irreversible action is inbound. Set your defenses, then brace: