Diverging from conventional multi-agent systems that emphasize cooperation, BlackSwanX is an adversarial intelligence engine where AI agents "fight" each other. This system integrates up to 200 "citizen AI agents" designed to argue, panic, and emotionally spiral, while a designated "BlackSwan Assassin" actively attempts to dismantle emerging consensus. A key differentiator for BlackSwanX is its 100% local operation on the Ollama platform, ensuring zero API costs and maximizing the exploration of chaotic dynamics for unique insights.
BlackSwanX features a diverse array of unique AI agent personas, including a "Vedic Astrologer," a "Panic Seller," a "Chaos Mathematician," a "Gen Z Culture Decoder," and a "Street Smart Hustler" (known for saying, "your pitch deck is pretty, show me your bank account"). These agents collectively endeavor to predict future outcomes through adversarial interactions. The system's primary goal isn't to elegantly solve specific real-world problems, but rather to pinpoint inaccuracies or blind spots within crowd consensus.
The BlackSwanX project is open-source on GitHub (project link provided in original), complete with a two-minute quick start guide. An example run illustrates its distinctive prediction format. When queried about "Will NVIDIA crash when the AI bubble pops?", the system generates a "Kill Shot" prediction (e.g., Quantum computing making GPUs obsolete, 10% probability), aggregated "Citizens'" sentiment (e.g., 25% bull / 65% bear), a "Dissonance" score (e.g., 33.6/100, indicating maximum chaos), and an "Antifragile Play" recommendation (e.g., Diversify into quantum computing partnerships).
BlackSwanX leverages 174 AI experts and 200 citizen agents, operating with zero API costs entirely on a user's laptop to predict various scenarios. Its foundational principle states: "Where the crowd is wrong, the alpha lives." Unlike conventional prediction tools that reflect collective opinion, BlackSwanX's objective is to expose the crowd's erroneous judgments. The system deliberately avoids seeking consensus; instead, it targets the widest possible "Cognitive Dissonance"—the significant disparity between mass belief and expert concerns. This identified gap is posited as the source of "alpha" or outsized returns.
BlackSwanX demonstrates considerable advantages when compared to other multi-agent systems. Regarding cost, BlackSwanX operates entirely free, leveraging Ollama, whereas comparable systems like BettaFish and MiroFish incur expenses from multiple API keys or cloud services. Setup time for BlackSwanX is remarkably brief at 2 minutes, significantly faster than the 15-30+ minutes required by alternatives. Furthermore, BlackSwanX integrates 174 domain expert agents and 200 citizen agents (operating via the Shadow Swarm framework), surpassing competitors that feature fewer expert agents or rely on generic personas and different simulation frameworks.
The core architecture of BlackSwanX is built upon three entirely local and free large language models:
- Swarm: Utilizes the
llama3.2:3bmodel to simulate debates among 200 biased citizen agents. - Assassin: Employs the
phi4:14bmodel, specifically for generating "kill shot" reasoning and critical analyses. - Nexus: Based on the
mistral-small:24bmodel, responsible for synthesizing information and constructing Directed Acyclic Graphs (DAGs) to coordinate system operations.