A God-mode simulator where thousands of LLM agents argue the future

What it does MiroFish ingests real-world seed material—news articles, policy drafts, financial signals, or even novel chapters—and spins up a high-fidelity parallel society populated by thousands of LLM-backed agents. Each agent carries its own persona, long-term memory, and behavioral logic, interacting in a simulated environment that evolves over time. You describe what you want to predict in natural language, inject variables mid-run from a “God’s-eye view,” and receive a detailed forecast plus an interactive world you can interrogate afterward.

The interesting bit Rather than treating prediction as a single-model regression problem, MiroFish frames it as emergent social dynamics: it uses the OASIS engine to let agents gossip, argue, and react, then captures the collective outcome. The README even suggests you can use it to generate the lost ending of Dream of the Red Chamber or stress-test a university PR crisis before it happens.

Key highlights

• Built on the open-source OASIS social-simulation engine from CAMEL-AI, with added GraphRAG memory construction and report generation.
• Supports dynamic “God’s-eye view” intervention: inject variables during a simulation to test alternative futures.
• Dual-platform parallel simulation with dynamic temporal memory updates and a post-sim chat interface with individual agents.
• Incubated by Shanda Group; demo includes live public-opinion and literary-what-if scenarios.
• Requires external LLM and Zep Cloud API keys; the docs recommend Alibaba’s Qwen-plus and warn that costs spike past ~40 rounds.

Caveats

• The simulation engine is powered by OASIS, so the core social-interaction layer is an upstream dependency rather than original infrastructure.
• Financial and political prediction demos are listed as “coming soon,” so those use cases remain aspirational.
• Token consumption scales with agent count and simulation rounds; the maintainers explicitly flag high cost and suggest keeping early tests under 40 rounds.

Verdict Worth exploring if you research multi-agent emergence, policy rehearsal, or narrative forecasting and can stomach the API bills. Skip it if you need deterministic, lightweight prediction without LLM overhead.