The demo in one GIF
The one prompt
The agent got a single instruction: build a Google Maps clone with pan, zoom, search and markers, and make it run locally. No starter code, no follow-up prompts, no hand-holding.
Everything after that — reading the plan, writing the files, starting the app and checking the result — happened inside the same run. The trace format is identical to the console’s Deep Dive view: every step, every tool call, every line of output.
What the agent did
- Planned the page structure: map canvas, controls, search box, marker list.
- Wrote the HTML, CSS and JavaScript as separate files.
- Started a local server and opened the app.
- Verified layout and interactions against its own checklist.
- Reported back with the finished app and the run summary.
The honest numbers
| Metric | Value |
|---|---|
| Model | deepseek-chat |
| Agent steps | 24 |
| Tokens | 76.5k |
| Cost | ≈ $0.04 |
Why this matters
This is one recorded run, not a benchmark and not a guarantee. But it shows the loop people buy an agent for: an idea in, a working app out, with the plan and the proof attached. See how it compares to what an AI agent actually is, and watch the same loop build an App Store clone or a chess game.
Run it yourself
Install the agent, open the console and paste your own one-liner. Start with the getting started guide, watch your Deep Dive traces, and keep an eye on cost with token budgets.