Developing a Natural‑Language Game with Amazon Q CLI: Mastering AI Agents — Part 1

This post chronicles how I completed a Python Pygame Tic‑Tac‑Toe game using the terminal‑based AI agent Amazon Q CLI with nothing more than concise prompts.

What Is an AI Agent?

An AI Agent is an autonomous system designed to reach a user‑defined goal by running a “Observe → Plan → Act” loop powered by an LLM. It can call external tools, access the file system, hit APIs, and continually refine its strategy based on feedback.

If a regular LLM chatbot is a knowledge engine that stops at “question → text answer,” an AI Agent is an execution engine that turns that knowledge into real‑world actions that change your system.

Put differently, the LLM is the brain; the agent adds hands and feet (tool usage and iterative autonomy) so it can finish larger goals on its own.

LLM vs. AI Agent Example

• LLM chatbot: “What’s the weather in Seoul?” → “It’s sunny today with a high of 25 °C.”

• AI Agent: “Build a simple web app that shows Seoul’s weather in real time.”

  1. Observe: Check current directory and installed libraries.
  2. Plan: Write weather.py → call OpenWeather API → set up a Flask server → create HTML template → craft run script.
  3. Act: pip install flask requests, generate code files, launch the local server.
  4. Feedback: Open the page, confirm data loads, tweak CSS for UI polish.

  The user typed a single sentence, but the agent completed the entire code‑build‑run cycle and produced a working app.

---

Observe‑Plan‑Act (OPA) Cycle

To move beyond text answers and actually edit code or deploy systems, an agent must decide what to do next on its own and adjust its behavior based on results. That mechanism is the Observe‑Plan‑Act (OPA) cycle.

Why OPA Matters

1. No observation → blind execution — it might touch the wrong file or revisit already‑fixed issues.
2. No plan → inefficiency — without an optimal route (e.g., UI → logic → tests) iterations multiply.
3. No feedback → unknown quality — a self‑correcting loop is essential for reliability.
4. With OPA → multi‑step workflow automation — even long pipelines like “code change → build → test → deploy” can finish from a single prompt.

Phase‑by‑Phase Flow

1. Observe — Gather the current state: prompt input, command output, code, logs.
2. Plan — Draft an action sequence to reach the goal (e.g., “add AI opponent to Tic‑Tac‑Toe”).
3. Act — Interact with external tools: shell commands, API calls, file edits.
4. Feedback — Re‑observe results, update the plan, and loop.

---

AI Agent Line‑Up

• Amazon Q CLI: Terminal‑focused development agent.
• GitHub Copilot Agent mode: Automates PR creation and refactoring tasks.
• OpenAI Codex: Software‑engineering agent.

Q CLI Installation & Usage

Amazon Q CLI is a terminal‑based AI agent that automates development tasks like code generation, refactoring, testing, and deployment through conversational prompts. It understands hundreds of CLI tools such as git, docker, and aws, and even suggests execution plans.

Installation is straightforward with the official installer — see the installation guide. For Windows you can run it inside WSL; refer to the Windows guide.

Because it must run inside WSL on Windows, the experience there can be a bit cumbersome.

On macOS, enabling Shell Integration unlocks handy features like inline command autocompletion.