Learnings and Findings from Building an AI Coding Agent - The Mini Cursor

In my recent project, I built an AI Coding Agent that acts like a “Mini Cursor” - a smart assistant that sits right inside your terminal or coding environment and helps you build software step by step, just like a human coding partner. This experience was eye - opening and gave me deep insights into how AI can seamlessly improve the developer experience.

What is an AI agent ?

Imagine you have a helpful assistant who listens to what you want and then takes action for you - like booking your tickets, finding the fastest route home, or even helping you write an email. An AI agent is just like that, but inside a computer.

Key Learnings -

1. Interactive Coding - Just like collaborating with a human teammate

Instead of just asking the AI to generate a block of code once and done, I realized how powerful it is to have a back & forth conversation with it. The agent can ask clarifying questions, break down big tasks, and improve the code step by step based on feedback.

2. Planning and Structure Matter

The AI agent works best when you treat it like a project partner who needs clear instructions. Breaking down your project into smaller, manageable tasks — like “build this React component” or “write this API endpoint” — helps the agent stay focused and deliver accurate results. Without good structure, the code can get messy or incomplete. So, being a project manager for your AI assistant is key.

When you ask the AI Agent to “Create a TODO app in React,” it doesn’t jump straight to code. It thinks step-by-step:

First, it plans:

{"step": "plan", "content": "User wants a TODO app in React"}
{"step": "plan", "content": "Ask user to choose a UI theme"}

Then, it takes action:

{"step": "action", "function": "select_theme"}

Then, it observes the result:

{"step": "observe", "output": "User selected 'PurpleLight'"}

And then the final output:

{"step": "output", "content": "React Todo app has been created with the PurpleLight theme and is running at http://localhost:5173. You can now access it in your browser."}

This structured loop - plan → action → observe - ensures that the Agent builds the project logically, one decision at a time, just like a human developer would.

Example -