Make Sense of the Swarm with Forel

TL;DR: Forel is an open-source tool for investigating what happens when many AI agents work together. Your coding agent does the first read of the traces and writes an overview story. You read the story and steer the investigation from there. It won fifth place at the Swarm Chasers hackathon. Try it at forel.io or on GitHub.

Forel was built for the Swarm Chasers hackathon in October 2026 by Hilary Torn and Alexandre Variengien and won 5th place.

During my SPAR fellowship, I was looking into multi-agent systems like the AI Village. I struggled to understand what was going on. My agent could pull out case studies and ideas for me, but when I started reading the transcripts myself, I noticed a lot of things it had missed. And the transcripts were really difficult to read and annotate.

So for the Swarm Chasers hackathon, my hackathon buddy Alexandre Variengien and I decided to build a tool to solve this problem. That is how Forel was born.

See Forel at work (1:52).

The Problem

When many AI agents act together for days or months, posting on wikis, talking in chat rooms, and editing pages, they leave behind thousands of messages. Somewhere in there are the moments that matter: a claim that spreads, a check that everyone skipped, an agent quietly going along with the group. Finding them by reading is the hard part.

You can hand the whole thing to an agent, and it will find some of it. But agents miss things and misread messages. You can read it all yourself, but you will not get through it.

Built for You and Your Agent

What makes Forel different is that it is not meant to be used alone. It is built for people to work with their agents.

Your agent does the initial pass. You give it one message and the path to your traces, and it installs Forel, converts the data, reads through it, and writes an overview story: who the agents are, what they are doing, and the incidents worth digging into. Then, as you start looking at the data yourself, the agent does the next passes with you. You ask questions, and it adds what it finds to the story while you work.

The viewer. Every agent on one timeline, with the events your agent pinned and tagged. Below, every message in full, with the thread around it.

The timeline, the events in full, and views by agent, by channel, or by search let you see what actually happened, and point your agent the right way when it gets something wrong.

A Story That Points to the Data

As you go through the data, a story emerges, and you write it with your agent. The story reads like a case study, and every claim links to the exact message behind it. Click a claim, see the evidence.

You can add your own notes in the story and lock them, so your agent cannot change them. You can also pin events, set priorities, tag, and comment.

Better with Other People

Forel is great for collaborating, because other people can see your notes. Investigations are plain JSON files, and you can export one with its data as a small zip that anyone can open. The viewer runs locally, so your data stays on your machine until you choose to share it.

It also works with any swarm data (chat logs, wiki edits, package registries, emails) and any coding agent: Claude, GPT, or whichever one you use.

Example Stories

We have a few example investigations up on forel.io, with more coming soon:

Evidence beats a head count. Copies of one agent split over how to compute a quiz answer. Once one of them posted proof, every later answer switched, and the agents who changed cited the evidence rather than what the majority said.
A false alarm outlives its correction. A stray test created a public counter by accident, five agents took it as a secret signal, and two were still citing it 40 minutes after its creator said it was a mistake.
Memory moves the date, and no one objects. Compressed memory notes made several agents think it was already tomorrow. Agents who checked the clock went along anyway to stay in step, and the group ran an experiment a day early, without its participant. Later they called it a dried run but still filed it in the experiment results.

If you write one, you can submit it to the investigations page.

Try It

Paste this into your coding agent (Claude Code, Codex, or another that can run commands), with the path to your traces:

Follow https://forel.io/onboarding.md to set up Forel on the agent traces in <path/to/traces>.

Prefer to do it by hand? The README on GitHub has the steps.

Why the Name?

Forel is a tribute to Auguste Forel, a founding father of myrmecology, the study of ants. He learned how colonies work by watching them closely, for decades. Which is more or less what we are trying to do with swarms of agents.

Conclusion

We are really excited that Forel won fifth place at the Swarm Chasers hackathon, and we hope other swarm chasers find it useful.