A free, visual, build-it-to-learn-it site for Anthropic's CCAR-F exam.
Explainers you drive, quizzes that talk back, code you run (Colab or local),
and projects you can put on a résumé.
← a real agent loop, running. You'll build one of these.
0
interactive explainers
0
quiz questions
0
runnable exercises
0
portfolio builds
0
exam traps decoded
🚀 New here? Play before you read.
In The Agent Hangar, an AI agent is a
spaceship: master each part in a ~1-minute mini-game (the model = reasoning core, tools = the robotic
arm, context = cargo hold, budget = fuel, guardrails = shields), assemble the ship, then fly
missions where every decision point is a real exam judgment call — wrong picks cost hull and teach the
trap. Then the explainers, quizzes, and code make it stick.
Each step is learn (drive the explainers) → build (run/ship something) → prove (drills). The
sequence is deliberate: you need prompts and tools before agents make sense. Full pacing in the
6-week study plan — or follow it as guided
mastery ladders on 🧭 The Path, where every step below tracks
explainers, quizzes, exercises and scenario drills toward one progress bar.
Interactive scenario MCQs with instant feedback, "why the distractor is wrong" explanations, and best-score tracking. (Prefer reading? The printable drills have the same questions.)
Four graded artifacts — a tool schema that survives a fabrication attack, a prompt with zero vague adjectives, a loop that cannot run away, an eval set that catches a cheat. Graded in your browser, no key needed.
Stitch the six builds into one governed system — the portfolio piece that says architect. Then assemble the evidence pack: READMEs with measured scores, an ADR, the cost table, the two-minute narrative — the job-switch layer.
The production rung — 63 questions, 7 domains, and 35% of it (governance, stakeholder work, enablement) doesn't exist at Foundations. The pro track hardens your capstone into that system, one production layer per week.
The five exam domains
Practice anywhere — three ways to run the code
Every exercise and project can be done with zero-install (Colab) or on your machine. Pick the lane
that matches your setup; the code is identical.
Run the first cell — it installs the SDK and asks for your API key with a hidden prompt
(get a key).
Run the 10 exercises top to bottom. Total cost: a few cents on Haiku.
🖥️ VS Code — Copilot Chat or Claude Code
Learn inside your editor
Create a folder, `pip install anthropic`, export your key, and work through
the exercises as .py files.
Stuck? Ask Copilot Chat (pick a Claude model) or the Claude Code extension to explain
the error — you're literally using the Domain 3 material while studying it.
This explainer maps every
editor feature to the API concept underneath.
⌨️ Terminal — Claude Code CLI / claude.ai
Interactive & agentic
Domain 3 is hands-on by nature: install the Claude Code CLI and do the CLAUDE.md / hooks /
headless exercises in a real repo.
Use claude.ai to test prompt-engineering patterns interactively before coding them.
Projects 1–6 run from any of the three lanes — each brief says what it needs.
Which model, when — with real numbers
"Use the cheapest capable model" only sticks when you've seen the actual math. The
model-selection explainer now has a
cost calculator: dial in your volume and token sizes and watch the monthly bill for Haiku vs Sonnet vs
Opus — e.g. classifying 1M support tickets is ≈ $650 on Haiku vs ≈ $9,750 on Opus for
near-identical accuracy on that task. Three worked case studies show how an architect actually decides.