# Tokyo Battle Env

- **Event:** [Gemini 3 ハッカソン 東京](https://cerebralvalley.ai/e/gemini-3-tokyo-hackathon)
- **When:** Sat, Feb 21 at 9:00 AM – 10:00 PM (GMT+9)
- **Where:** 東京都渋谷区渋谷 3 丁目 21 - 3 渋谷ストリーム 5F
- **Team:** [Soki Fukuda](https://cerebralvalley.ai/u/soki)
- **GitHub:** https://github.com/createcentury/Gemini-3-Tokyo-Hackathon/tree/main
- **Demo video:** https://youtu.be/pu-oF9JJRX0
- **Gallery:** https://cerebralvalley.ai/e/gemini-3-tokyo-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/gemini-3-tokyo-hackathon/hackathon/gallery/61

Real-Data Environment: Unlike synthetic AI benchmarks, Tokyo Risk grounds every experiment in reality. Google Places API feeds Tokyo's 23 wards , while Google Routes API injects real traffic conditions as dynamic combat penalties. Agents reason about a city that actually exists.
Autonomous Online Learning Loop: Each player deploys 10 agents configured with custom system prompts, tool selections, and multi-agent coordination strategies. After each battle, results feed back into an automated optimization cycle—Gemini analyzes outcomes, generates improved configurations, and reruns the experiment automatically. The system answers: "What prompt, tool, and agent strategy actually wins on real Tokyo geography?" Voice-Driven Sub-Commander: A sub-commander layer bridges human intent and agent execution via voice. Speak a command and Gemini parses it through Function Calling into live agent directives.

---

Markdown version of https://cerebralvalley.ai/e/gemini-3-tokyo-hackathon/hackathon/gallery/61. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt