# Ascent Performance

- **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:** [Mohamed Abdelnasser](https://cerebralvalley.ai/u/Mohamed7)
- **GitHub:** https://github.com/mohamedabdelnasser7-gif/unfenced
- **Demo video:** https://youtu.be/XkvQOXEDNX4
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4772

Unfenced is an engine that builds AI agents, trains them against
  realistic customers, and keeps them learning after they go live.

  You describe the agent you want in plain words. Unfenced writes its
  card (what it does, what a good and bad result look like, what it
  must never do), builds it, and lets you add your own data. Before it
  meets a real person, it practises in the Arena against agents made
  from real customer behaviour, not polite AI talk. Up to 10 matches
  play at once, and you watch every chat live, won or lost. If you
  don't like a reply, you correct it on the spot. The correction takes
  a short exam before the next match uses it.

  An agent goes live only after it passes a gate set by risk: 7 of 10
  wins for most jobs, and 16 of 20 with zero safety failures for
  high-risk fields like money or customer data. After that, every real
  case flows back into the Arena. Each night the agent studies its
  losses and writes its own lessons, and every lesson passes the same
  exam before it is used. Agents in the same field learn from each
  other's mistakes, so each new agent makes the whole engine smarter.

  The first agent is Sales for a football boot store on WhatsApp,
  replying in Egyptian Arabic, English and Franco. Any agent you build
  runs on the same engine. Built on Meta's Muse models.

## More from Meta Global AI Developer Hackathon

- [CoalitionKit](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4769)
- [Tercero](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4770)
- [Teach-Back](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4771)
- [Atlas Codes](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4773)
- [Oneiron](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4774)
- [One Voice](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4775)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4772. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt