A few weeks ago, Matías Molinas and Ismael Faro were talking about what should come next for Evolving Agents Labs—and how to focus our energy so the community around agent research is not scattered across a dozen small repositories, but working toward something larger together.
That conversation became a decision: Evolving Agents Labs is joining ANFAIA.
Evolving Agents Labs has been a place to ask difficult questions in the open. How do agents evolve without losing our ability to trust them? How do we notice when an agent's memory has been quietly poisoned? How do we make a small model incapable of producing invalid output, instead of merely asking it not to?
Projects including sleep-harness, agentvcs, skillos, token-trie, evolving-memory, and evolving-robot will continue as part of the ANFAIA community, alongside the rest of our experiments in adaptive agent systems. The full catalog lives at github.com/EvolvingAgentsLabs, with each project labelled honestly as Reproducible, Results, or Prototype.
Why ANFAIA?
Because ANFAIA is about connecting young talent with real open-source AI projects. And that is exactly what this work needs: more hands, more perspectives, and more people willing to publish the experiments that fail—not only the ones that work.
This is not a shutdown. It is a shared purpose.
The experiments will continue in the open, following the same principle we have tried to keep from day one:
Reproducible or it didn't happen.
Matías, I, and everyone who has contributed to Evolving Agents Labs so far will keep exploring these questions—now within a broader community built around learning, experimentation, and sharing what we find.
What's next: lora-kernel
The clearest view of where this is going is on the lab's site, evolvingagentslabs.github.io, and it centers on our newest and most ambitious proposal: lora-kernel.
The idea is simple to describe and hard to build. Instead of a multi-agent system assembled from a router model, JSON schemas packed into a system prompt, and a parser trying to infer whether a tool call was sincere, the whole agentic system becomes a set of QLoRA adapters over one resident base model. One GPU. One base model. A pool of small weight deltas that vLLM swaps per request. The framework does not get simpler by adding more scaffolding—it gets simpler by turning into weights.
We are being explicit about the pieces rather than hand-waving them away:
Right now, the label is honest: specified, nothing built. Every external claim is marked [read] and cited; there is no [ran] yet. That is the same discipline the rest of the lab follows, applied to the most ambitious project we have proposed so far. It is also a good place to start if you want to find the open questions sharp enough to contribute to.
The spirit carries forward
I hope this becomes a small working example of what grassroots collaboration in AI can look like—where research is not just something you consume, but something you can take part in.
Thank you to everyone who has contributed to Evolving Agents Labs so far. The spirit of the lab—curiosity, rigor, and publishing failures alongside wins—carries forward inside ANFAIA.
And this is only the beginning. More on what we are building next soon.
If agent memory, self-evolving systems, verifiable AI behavior, or building AI in the open is your thing—whether you are a contributor, mentor, or simply curious about these questions—let's talk.
Explore the projects: github.com/EvolvingAgentsLabs · See where it's headed: evolvingagentslabs.github.io