MCP-powered web agent
An exploration of how an agent can gather context, choose tools, and make useful progress while keeping a person in the loop.
Live demo coming soon
01 / The idea
What I'm exploring
The interesting part of an agent is not the model call. It is the context around the model: what it can see, what it can do, and how clearly it can explain what happened.
This project is a place to test those boundaries with small, observable workflows instead of pretending the whole system is magic.
02 / The approach
Small steps, visible trade-offs.
- Start with a narrow task and a small set of tools.
- Make every action inspectable and easy to interrupt.
- Use the results to improve the interface before adding more capability.
03 / The outcome
What success looks like.
The goal is a reliable experiment that demonstrates useful AI interaction without hiding the reasoning or the trade-offs.