# Agentztein

- **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:** [Alan Jesus Lopez Jacinto](https://cerebralvalley.ai/u/alan-jacinto)
- **GitHub:** https://github.com/AlanJLJ2202/Agentztein
- **Demo video:** https://youtu.be/G52-odMUHcM?si=xOIigUd8Z3WZVGpW
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3674

Agentztein - Self-Evolving AI Agents with Verifiable Improvement

What if AI agents could improve themselves without losing what they already know?

Agentztein is an experimental self-evolving AI agent powered by Meta's Muse models, built to investigate one of the most important challenges in autonomous AI: how can an agent continuously improve its capabilities without introducing regressions?

Unlike traditional agent frameworks that focus primarily on task execution, Agentztein focuses on measurable, testable, and reversible self-improvement.

The system enables agents to identify weaknesses, propose context-specific skills, evaluate improvements against existing capabilities, and preserve changes only when they pass explicit validation criteria. A deterministic evaluation harness measures performance across different contexts, helping detect when an improvement in one area negatively affects another.

Agentztein also explores self-modifying code through sandboxed execution, versioned agent cores, automated testing, and controlled activation. Its memory consolidation system captures lessons from previous experiments, while interactive 3D visualizations make agent execution, decisions, costs, and evolution observable.

Powered by Meta's Muse Spark and Muse Wonder, Agentztein combines autonomous planning, experimentation, regression testing, human oversight, and rollback mechanisms within a unified research environment.

**Our vision is to move beyond AI agents that simply execute tasks toward agents that can evolve responsibly, demonstrate measurable improvement, and provide evidence that their changes actually work.**

Self-improvement is easy to claim. Hard to prove.

## More from Meta Global AI Developer Hackathon

- [Zazen Builder](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3671)
- [Italian Rides](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3672)
- [Medha](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3673)
- [Team Juicy](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3675)
- [DeepKomal](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3676)
- [Minkyu Park](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3677)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3674. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt