MiroFish, an AI prediction engine that simulates thousands of autonomous agents in digital worlds to forecast outcomes, rocketed to #1 on GitHub Trending on March 7, 2026. The project has accumulated 28,600 stars, with 2,782 gained in just the last 24 hours. Built by Guo Hangjiang, a 20-year-old undergraduate at Beijing University of Posts and Telecommunications, in 10 days, the project secured a 30 million RMB ($4.1M USD) investment from Shanda Group founder Chen Tianqiao within 24 hours of demo submission. Moreover, it surpassed repositories from OpenAI, Google, and Microsoft on its way to the top spot.
Unlike traditional machine learning that trains models on historical patterns, MiroFish creates “digital petri dishes” where LLM-powered agents with distinct personalities, memories, and behaviors interact to reveal predictions through emergent social dynamics. Consequently, it’s not just another GitHub trend—early adopters are already deploying it for real business forecasting.
Digital Petri Dishes Replace Training Data
MiroFish flips the forecasting script entirely. Traditional ML asks “what patterns does the data show?” while MiroFish asks “what happens when thousands of realistic agents encounter this situation?” Users feed “seed” information—news articles, policy drafts, financial signals—into the system, which constructs high-fidelity parallel digital worlds populated by autonomous AI agents. Each agent has independent personality profiles, long-term memory systems, and behavioral logic powered by LLM reasoning.
Built on OASIS (Open Agent Social Interaction Simulations) from CAMEL-AI, the framework is designed to scale up to one million agents. Furthermore, agents can perform 23 different actions—following, commenting, reposting—and maintain memory that evolves over simulation time. The key innovation: predictions emerge from agent interactions rather than pattern matching. Additionally, users can inject new variables mid-simulation from a “God’s-eye view” to test “what if” scenarios and directly query agents to understand their decision-making. This interpretability is something black-box ML models fundamentally can’t provide.
From Public Opinion to Financial Markets
The platform’s versatility is its strength. Public opinion simulations can predict how sentiment evolves after news events, generating 90-day trajectories that show polarization curves and media amplification cycles. Financial forecasting captures trader psychology and herding behavior that traditional models miss—simulating differential reactions from retail investors versus institutional funds to the same market event. Similarly, policy impact analysis reveals how different stakeholder groups respond, form alliances, or organize resistance before implementation.
The team even demonstrated creative applications: they fed MiroFish the first 80 chapters of a classic Chinese novel and had it predict the lost ending based on character behavior patterns. Organizations like “The Zero-Human Company” have publicly announced adopting the platform for business forecasting, describing it as a “digital crystal ball” for scenario planning that outpaces traditional analytics. Use cases span from testing how university controversies might unfold across social media to war-gaming competitive responses to product launches.
The 20-Year-Old Who Built It in 10 Days
Guo Hangjiang’s story challenges the conventional startup playbook completely. Using “Vibe Coding”—rapid prototyping with AI coding assistants—he built MiroFish in 10 days and recorded a simple demo video. He sent it to billionaire Chen Tianqiao’s desk, and within 24 hours, Chen committed $4.1M to incubate the project.
This wasn’t Guo’s first viral hit, however. His previous project BettaFish, a sentiment analysis tool, gained 20,000 stars on GitHub in a week. Chen Tianqiao has been promoting a “super-individual” theory: that in the AI age, one person can accomplish what previously required an entire company. Therefore, MiroFish embodies this vision—a solo undergraduate building a platform that challenges enterprise forecasting systems.
The GitHub metrics tell the story: 18,000 stars and 1,900 forks within days of launch, surpassing projects from OpenAI, Google, and Microsoft on the trending list. As a result, the rapid institutional backing validates that the industry sees significant potential in this approach, not just developer curiosity.
Simulation vs. Traditional ML: When to Choose
Skip the hype—MiroFish comes with serious trade-offs. Computational costs are high. Each agent requires LLM API calls, so simulations with thousands of agents generate thousands of inference requests. The project is early-stage (V0.1.2, only 4 months old since November 2025) with ongoing development. Validation is challenging since verifying prediction accuracy requires waiting for real-world outcomes. Unlike traditional ML which produces deterministic results, agent-based simulations include stochastic elements that can produce different outcomes for identical scenarios.
Research on multi-agent forecasting indicates these approaches “achieve a level of forecasting precision previously unattainable” while providing “continuous learning and adaptation advantages,” but at computational expense. Nevertheless, early adopters report the steepest learning curve is understanding when simulation-based forecasting adds value versus when traditional ML is sufficient.
Smart teams will use hybrid approaches: traditional ML for pattern-matching problems (financial time series, demand forecasting), MiroFish for complex social dynamics and scenario exploration (opinion evolution, policy impact, behavioral forecasting). Indeed, this isn’t about replacement—it’s about expanding the forecasting toolkit.
The Future of Forecasting
MiroFish signals a broader shift from model-centric to simulation-centric AI. As LLM APIs become cheaper and agentic AI frameworks mature (2026 is seeing an explosion of agent systems), multi-agent simulation becomes economically viable for practical forecasting. Notably, the OASIS framework MiroFish builds on was only open-sourced in December 2024, yet by March 2026 it’s powering a #1 GitHub trending project with commercial backing.
Whether MiroFish specifically dominates or competitors emerge, simulation-driven forecasting using multi-agent systems is here to stay. Teams building forecasting, scenario planning, sentiment analysis, or decision support tools should experiment with these approaches now to understand their strengths and limitations before they become industry standard.
Key Takeaways
- MiroFish represents a shift from data-driven to simulation-driven forecasting through agent interaction
- Handles social dynamics and emergent behavior that traditional ML pattern-matching misses
- Trade-off: Higher compute costs but interpretable and scenario-exploration capable
- Early adopters already deploying for public opinion, financial, and policy forecasting
- Signals broader trend toward multi-agent simulation as complement to traditional ML
