What happens when you stop telling an AI exactly what to code, and just give it a goal?

In this video, I give Pi Agent a single prompt to build a landing page from scratch, then let it take over the workflow. It breaks the task down, creates the files, checks its own work, fixes an error, and later adds Chinese language support - with surprisingly little input from me.

The website is just the experiment. What I'm really interested in is the workflow behind it: the space between giving an AI a goal and getting a finished result.

For this setup, I'm running Pi with Qwen models locally through llama.cpp, switching from Qwen3.6-35B-A3B to Qwen3.5-9B for the lighter coding tasks.

This isn't a benchmark or a tutorial. It's me trying to understand what actually changes when an AI goes from answering prompts to working through a task on its own.