MiroFish: Deploy a Swarm of AI Agents to Build Knowledge Graphs and Predict the Future

Key Takeaways

• MiroFish runs hundreds of agents in parallel across hundreds of iterations, each exploring different reasoning paths through a prediction problem. The swarm isn't just redundancy — divergent agents force the system to surface edge cases hidden from any single-pass analysis.
• The output is a knowledge graph, not a single answer: a visual map of relationships, confidence levels, conflicting conclusions, and consensus points across the entire agent swarm. This is richer than any point estimate.
• Swarm intelligence emerges from disagreement: agents that reach conflicting conclusions are the most valuable signal. The system tracks where consensus breaks down to highlight the highest-uncertainty regions of the prediction.
• Dubai real estate demo: the knowledge graph revealed a non-obvious correlation between EXPO legacy infrastructure and price trajectories in specific districts — a finding that emerged from cross-agent comparison, not from any individual agent's analysis.
• Cost and hardware: running hundreds of API calls per simulation makes this expensive at cloud rates. Local models (Ollama) dramatically reduce cost but require 16GB+ VRAM for the model quality needed for coherent reasoning chains.

How MiroFish Works

The architecture is straightforward: a controller spawns N agents with the same question but different reasoning seeds (temperature, context framing, starting assumptions). Each agent runs independently through the problem, building its own reasoning chain. After all agents complete, a synthesis step aggregates their outputs into a knowledge graph — nodes are concepts, edges are relationships, and edge weights reflect cross-agent consensus strength.

The knowledge graph is interactive: zoom in on any node to see which agents contributed to it, what conclusions they reached, and where they diverged. High-divergence nodes are the most interesting — they represent the genuine uncertainties in the prediction, not just noise.

Tim runs the demo on sample data he provides, showing the full pipeline from data input to final graph output. The tool is open-source and can be run locally with an Ollama backend or against cloud APIs.

Related on OpenClawDatabase

• Hermes + OpenClaw Multi-Agent Setup — practical multi-agent architecture for production use
• OpenClaw + Hermes Workflows — four patterns for coordinating multiple agents
• OpenClaw Hub — multi-agent capabilities and configuration