Chaos Lab

**Judgement:** A strong, opinionated research/teaching tool for illustrating AI alignment and multi-agent value conflicts using Gemini models. Best suited for labs, educators, and safety researchers; not a general-purpose automation skill. **What it does well:** - Spawns multiple Gemini agents with conflicting objectives and logs their behavior in a shared sandbox. - Provides ready-made experiments (duo and trio) plus scripts to compare different Gemini models. - Offers clear customization points (agent prompts, sandbox contents) and good documentation with examples. **Main risks & limitations:** - **External dependency:** Entirely depends on Gemini APIs; experiments break if API access, quotas, or model names change. - **Data exposure:** Any workspace content you test is sent to Gemini; not appropriate for sensitive or proprietary data without additional controls. - **Non-determinism:** Results are stochastic and model-specific; not suitable for reproducible, production-grade workflows. - **Scope creep:** It is a research/demo framework, not a safety guarantee mechanism; using it as “evidence” of model safety or robustness would be misleading. **Recommended scenarios:** - University or industry workshops on AI safety and alignment. - Internal labs wanting quick, vivid demonstrations of misaligned objectives in multi-agent systems. - Prompt engineers and researchers comparing model behaviors and personas under conflicting value setups. **Not recommended for:** - Production decision-making or automated file/system management. - Any environment where you cannot safely send sandbox data to an external API provider. - Use cases requiring stable, deterministic outputs or strong security assurances beyond the documented sandboxing.