# Agent Boost

- **Event:** [Gemini 4 NYC Hackathon](https://cerebralvalley.ai/e/gemini-4-nyc)
- **When:** Sat, Nov 21 at 9:00 AM – 10:00 PM (EST)
- **Where:** New York, NY
- **Team:** [Michi Yamamoto](https://cerebralvalley.ai/u/michiyamamoto)
- **GitHub:** https://github.com/michi883/agent-boost
- **Demo video:** https://www.youtube.com/watch?v=9lhYcQlgfUY
- **Gallery:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/42

Most AI agents are limited by the tools they already have. I wanted to see if one AI could build new tools for another AI on demand.

That's the idea behind Agent Boost.

Gemma 31B runs locally on my Mac and creates a basic slide. I then ask Gemini to add a new capability, like interactive 3D visualization. Gemini builds and tests the tool, and Gemma discovers and uses it on its own.

To see if this works beyond one example, I tested it on four tasks, including three unseen requests. Gemma alone succeeded in 0 out of 4. With the Gemini-built tool, it passed all 4.

The tool itself takes about 15 seconds to build, using Gemini Decisions, Hydrogen, and Neon.

## More from Gemini 4 NYC Hackathon

- [FightRoom UFC](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/39)
- [heliocam](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/40)
- [Pour Labs](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/41)
- [@labs](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/43)
- [Dunkoma](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/44)
- [INTERLOCK](https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/45)

---

Markdown version of https://cerebralvalley.ai/e/gemini-4-nyc/hackathon/gallery/42. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt
