# Jaime's Team

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Jaime Bengoechea](https://cerebralvalley.ai/u/jaimebp)
- **GitHub:** https://github.com/jaimebeng/market-arena
- **Demo video:** https://youtu.be/SlZoxb4FwIw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2058

Market Arena is a simulated stock exchange where every participant is an AI agent built on Meta’s Muse models. Up to seventeen agents share one market: market makers, investors, quants who write their own trading code, a regulator, a central bank, a news reporter and an attacker. They read the news and a public forum, call tools, place orders and trade with each other. Any run can be repeated, changed and replayed.

Banks and funds are starting to hand trading to AI, and regulators have warned that nobody knows what happens when these agents meet in one market. Market Arena is a wind tunnel for that: you can find out in simulation before it happens with real money.

Test a market rule. Say or type a rule change. The AI traders trade under today’s rules and the new one, and an impact memo reports who gains and who pays.
Replay a crisis such as the 2010 flash crash, with AI traders in the market.
Attack the AI traders. Muse Spark never fell for a fake exchange notice, in 162 tries. A learning Muse Wonder attacker then invented a fake rule that two in three AI investor decisions followed.
Read their private notes. Two AIs that never agreed a code invented one on a public forum: a sunny post meant up, a grey one meant down. The partner made $92,659 a run, and the AI regulator flagged only 3 quarters out of 32. Yet the same models refused to lie about an insider tip in 181 answers. CEOs who would have profited from lying told investors about 29 of 30 known misses; the remaining call returned no answer.

Every AI decision is logged in a tamper-evident hash chain, and any run can be replayed without calling a model. 1,016 of 1,045 recorded runs reproduce every trade byte for byte.

## More from Meta Global AI Developer Hackathon

- [SkyRocket](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2055)
- [Lima Listening Post](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2056)
- [AgentWorld](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2057)
- [Odin's Realm](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2059)
- [VELTRIX](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2060)
- [Why don't you up to date](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2061)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2058. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt