Deploy an AI Agent Swarm to Analyse Anything - MiroFish Research Walkthrough!

MiroFish Research has had some big updates, so in this video I walk through the whole thing from install to final report. You spin it up on your own machine, feed it news, let it predict possible events, then watch a swarm of AI personas argue it out until they reach a consensus. All running locally with Ollama!

How do you install it?

It's all driven by a Makefile. make install sets everything up, including spinning up the Postgres database and creating the tables from the schema. Then make start launches the app.

I've put some testing scripts into the make commands to check that Ollama has a model loaded and is actually responding to a chat request, so you're not left guessing why nothing works. Once it says all services are loaded, click the button and your browser opens on localhost port 3010. The first launch took around 20 or 30 seconds for me.

The full step by step install guide lives in my Skool at liberati.life.

What data goes in?

The dashboard pulls a lot of news from different feeds, some through MCP servers and some through RSS, and puts it into a format the AI can understand. In the demo there are 107 articles coming in, filtered by topic up the top. Right now that's oil and stocks, and the plan is to let you type any topic you like. I want to track things like military spending budgets and inflation rates around the world, and have it go off, pull in the news and start graphing it.

Each topic can be viewed four ways:

View	What it shows
Map	Where in the world the news is coming from
Chart	Price feeds over time
Table	Data like crude oil prices in rows
Articles	The news itself, with images

The left column holds the news feed, and it isn't just recent stuff. It also pulls historical news going back a year, two years or even ten, to give the AI the wider macro view. If you're looking at oil, what happened during previous wars in the Middle East matters, and I find that context is really important to hand to the AI.

How does the prediction work?

From the current and historical news, the AI generates possible events that might happen in the near future. To be clear, these haven't happened, they're predictions to analyse. I picked the most topical one, "Iran war escalates to full-scale invasion", meaning US ground forces actually deploying onto Iranian territory.

The layout is data feeds and prediction events on the left, and the MiroFish and Oasis side of the house on the right. Both columns slide out (I did some UX fixes there too). Once the event is selected, you hit Run Analysis and four AI-generated personas get to work.

For the demo I hooked it up to Qwen 3.5 0.8B, which is very small and not the smartest model in the family. You really should ramp it up to the 4B or 9B for better results. If you'd rather use a cloud model, drop a Gemini, Anthropic or ChatGPT key into the env file and those show up as options too.

What happens when the agents debate?

This is the fun bit. In round one each persona, such as an oil market trader, writes its own analysis with a call, its top three risk factors and recommended positions. Then they debate each other's answers, doing about three rounds of debate and rebuttals, trying to get to a better result. At the end they reach consensus and produce a final report, with each persona writing a summary of what it found.

The consensus price came out at $364 a barrel, which is a bit crazy. The oil market trader said $125, but the consumer impact economist blew the average out by going to $1,200. I don't think it's going that high, that would be absolutely crazy for the whole world economy. Honestly though, that's what makes it interesting. You could run another 10 agents with different job roles and perspectives from around the world and see how the picture changes.

So what's next?

MiroFish Research is my personal demo to prove it works for me. I'm using it for a new YouTube channel about macroeconomics and geopolitics based on real data feeds that I try to keep non-biased. I don't just build things, I actually use them, and if something doesn't work I throw it out. So far this one is doing really well!

The whole source code is shared in the Liberati community. Jump in at liberati.life and get your own swarm running. Cheers!