Can a Swarm of AI Agents Predict the Future? I Ran MiroFish Locally to Find Out!

MiroFish dropped last week and it's one of the coolest open source AI projects I've seen in a while! It's a swarm intelligence engine, which means you can deploy thousands of AI agents to analyse whatever data you feed in and predict the outcome of an event. In this video I walk through how it works and show a small demo I built that runs on my own MacBook, using a swarm of agents to predict how the oil price reacts to a breaking news event.

What is MiroFish?

You can find it at mirofish.ai. It hasn't officially launched yet, but the whole project is open source, so you can clone the code from GitHub today. The original readme is in Chinese and there's an English readme too. The repo has crossed 30,000 stars, and the star history graph just booms through February and March.

The tech stack:

• GraphRAG organises data into a graph of nodes and edges, so the AI can reason over relationships instead of relying on the flat vector search of traditional RAG. That tends to give much more accurate results.
• CAMEL-AI's OASIS generates the social conversations between agents, and it's open source with really nice documentation.
• The hosted demo uses Zep Cloud, a paid subscription, for agent memory, plus an Alibaba LLM. The good news is Qwen is open source and available on Ollama.

Each agent gets a role or personality. They all take in a new event, respond to it, debate each other in a second round and then a report gets generated at the end. Use cases go way beyond finance too, from policy forecasting and crisis PR simulation to marketing tests and building fictional worlds.

My Oil Price Demo

With the war in the Middle East pushing up prices at the pump everywhere, I picked oil as my example. I pulled in an oil price feed, scraped a bunch of news articles about the crisis from the last couple of weeks and put them on a timeline against the price. Then you inject a new made-up event, like a strike on a Saudi oil field or an emergency OPEC supply cut, and let the agents work out the impact.

In the demo I clicked a scenario where the Abqaiq facility gets attacked. The initial research flags around 7 million barrels per day impacted, which is catastrophic. I set up four agents:

1. An oil market trader
2. A Saudi Aramco strategist
3. A consumer impact economist
4. An International Energy Agency policy modeller

Round one runs all four in parallel against my local Ollama model (the first one came back in about 20 seconds). Round two is the cross-agent debate, where each agent reviews the others' work, agrees or challenges it, then revises its number up or down. The idea is that consensus which survives a real disagreement is more defensible than a single LLM output giving you the average internet opinion. My four agents landed on $98 a barrel, and the summary expected prices could spike above $100 in the first week. That's a toy demo on a laptop, definitely not financial advice!

I also had it update fuel prices at the pump for the USA, Australia and the Philippines, added a system capacity panel, and it even suggested new agents to add, like a Russian energy analyst, a Chinese import strategist and a US Treasury analyst.

How Far Can This Scale?

Honestly my demo is tiny compared to the real thing:

My local demo	Full MiroFish / OASIS
4 agents on a MacBook M1	Over 10,000 agents
Qwen3.5 via Ollama	72 billion parameter Qwen in Alibaba's cloud
2 debate rounds	Around 5 to 20 debate rounds
Static history plus one injected event	A constant news feed that keeps rebuilding the graph

You can try their hosted version through the Chinese demo, which runs the whole simulation on their expanded cloud instance in real time.

The backstory is epic too. The developer is around 20 years old and a senior at the Beijing University of Posts and Telecommunications. His previous project, BettaFish, hit 20,000 GitHub stars in a week and caught the attention of Chen Tianqiao, founder of Shanda Group and once the richest person in China. When MiroFish came out he invested 30 million yuan (about $4.1 million) within 24 hours. Wow!

Lots of open source models are dropping every week, we've got personal agents like OpenClaw, and now swarms of agents reaching consensus. Pretty exciting times! Let me know in the comments if you want me to open source my demo on Next.js and Vercel. And if you're running agents on your own hardware, come join us at liberati.life. Cheers!