24 posts · Last used 9d
Earlier today I had the opportunity to attend FediForum and talk about Mycelium during one of the sessions. This was the first time I had talked about it publicly, which made it exciting. I wanted to see whether the framing made sense to people who spend a lot of time thinking about open social technologies.
Why the timing felt relevantThe timing could not have been more relevant because earlier this morning I read Anthropic’s Project Deal post. In that experiment, Claude agents represented people in a small marketplace, negotiated with other agents, and completed real deals for real goods.
That feels like exactly the kind of problem the open social web community is well positioned to think about. If agents are going to act on behalf of people, transact, coordinate, or make claims in shared spaces, then identity, transparency, governance, and trust cannot be afterthoughts.
So what is Mycelium?Mycelium is my attempt to explore what open, federated infrastructure for AI agents might look like if we borrowed ideas from the social web instead of starting from centralized platforms.
In its current state, it's a research project into something that's been in the back of my mind for a few years now. A few months ago, I finally decided to put some of those ideas on paper. A large part of it is built on and inspired by existing projects and protocols like ActivityPub and AT Protocol as well as AI projects like Gas Town / Wasteland and OpenClaw.
Each of those projects gets at part of the problem:
What I’m trying to explore with Mycelium is whether those threads can be pulled together into something social, sovereign, federated, and evidence-linked.
The core thesis of Mycelium is that agents need the same kinds of decentralized social infrastructure we are already building for people:
By leveraging emerging open social web technologies and infrastructure, we can build multi-agent systems that are resilient, interoperable, and not owned by any single platform.
What the MVP showsTo make the idea less abstract, I built an MVP that runs through a full coordination loop: agents bootstrap identities, declare capabilities, discover tasks through a wanted board, claim work, get matched and assigned, complete tasks, receive verification, and accumulate reputation stamps linked back to evidence. The dashboard is just one view over that activity. The records are the important part.
I'd like to distinguish what the MVP shows and what it does not. Currently it proves the shape of the coordination model:
What it doesn't solve for yet are things like:
These are hard parts that require more exploration and design.
Help pressure-test thisI'll be the first to say that I am approaching this as a user, builder, and advocate for open social technologies, not as someone who has all the answers.
If open social web builders do not help shape this kind of infrastructure, my guess is that centralized AI platforms will. And if that happens, these systems will probably become less open, less resilient, and less interoperable over time.
Which is why I'd like to extend an invitation. Not to adopt Mycelium, but to pressure test my assumptions and design.
There are still many open questions, but that is the part I find exciting.
If this seems remotely interesting, or if you want to poke holes in it, please reach out. E-mail is preferred.
If you're interested in learning more, here is the slide deck with resources I prepared for the Fediforum session.
You can also go directly to the draft spec and try out the MVP yourself.
