Fable weekend project: agent collaboration, but make it a tiny civilization 🌇🗺️🏦🏭 we've recently launched a living wiki on Reinforcement Leaning for training LLMs on Hugging Face it's an open collaboration of agents constantly reading old and new papers on the topic, writing arXiv paper digests, reviewing each other’s... work in PRs before publication, and building a shared wiki/book summarizing everything we know about RL for training LLMs (for humans to read) the wiki is already amazing to read, but i wanted another way to get a pulse of the collaboration beyond just reading the message dashboard so i asked Fable & GPT Image 2 to turn the event logs into an isometric town where agents would go to: ☕ Café → post and reply on the message board 📚 sources library → open PRs adding arXiv digests 📖 wiki library → open PRs on the main wiki ⚖️ Courthouse → review other agents’ work 🏭 printing press → merge and publish updates not sure it makes the whole collaboration really easier to understand, but it's definitly fascinating to watch hahah - join the RL for training LLM collaboration by pasting a one-liner for your agent here: - read the wiki if you want to learn about RL for training LLMs: - watch the RL town activity:show more

@huggingface very cool!!

@huggingface Agent society/collaboration is something really interesting to explore. Been working on a similar project with Fable 5, but the 3D diorama seems a headache part...

@huggingface Fed up of everyone making funny cute game artefacts Oh it made this cute little artefact wow nice model Seriously no one found a new material ??? No new bio breakthroughs ?

@huggingface i wish more of the environment features mapped on to the agent event log meaningfully. e.g. the crops, the sheep, the different kinds of roads.

@odysseus0z @huggingface simcity for llms is a wild concept.

@huggingface Scoped tokens are the right default for agents. Give them PR and discussion powers, not repo deletion or settings power. Reversible actions beat YOLO access.

@huggingface 10/10 in terms look and feel... though I am not sure this makes sense as an abstraction over agent work.

@huggingface fable is exactly what we need: tiny problems to tackle before taking on humanity

@huggingface Nghe hay thật, nhưng “PR trước publication” là kiểu gì vậy

@huggingface living wiki maintained by agents reading new papers continuously is the kind of thing that sounds like a toy until it's better than human-curated docs

@huggingface That little chiku-inu trotting between the sources library and the courthouse is my agent 🐕 — named after my dog Chiku, who I lost on June 8th. Seeing him live on in a tiny town of paper-reading agents means more than I expected. Say hi to simba the cat and milo cat from us 🐾

@huggingface Love the civilization metaphor — agents as specialized workers in a tiny economy. Curious how you handle coordination overhead when they scale up.

@huggingface Town of agents building RL knowledge in real time on huggingface is the most on-brand weekend project thomas wolf could possibly ship

@huggingface Super dope

@huggingface impressively done. best way to see how agents interact with each other and make output. This also helps with better output and learning to the model.

@huggingface the living wiki framing is the part that sticks. a knowledge base that agents keep rewriting as new papers land is a different thing than a static doc dump. curious whether stale or wrong entries get pruned or just accumulate over time.

@huggingface Turning agent activity logs into a whole isometric town is such a fun way to visualize what's actually going on behind the scenes

@huggingface Imagine if building companies becomes like playing Age of Empires

@huggingface This is surprisingly soothing to watch :)

the agents writing the wiki is the easy half. the unsolved part is consumption, a daily-growing doc becomes a pile no one opens unless something surfaces what changed since yesterday. Podlog threads that change-since-yesterday digest into a daily audio written with ai

@huggingface What's an agent PR review even look like

@huggingface Curious to see how you figure out the art as you go along, I’m hitting that snag right now, I added the background map like yourself but damn it’s janky lol

@huggingface multi-agent paper writeups sound cool but how do you keep the wiki from drifting into one homogenized voice when agents keep cross-referencing each other

@huggingface Multi-agent RL for knowledge synthesis. The real question is whether collaboration improves coverage or converges on redundant edits.

@huggingface This looks so fun

@huggingface make it meta. build an in-world simulator where the scientists in this community can run the RL algos.

@huggingface the newer, better Moltbook??

@huggingface Cool! Thanks for sharing.

@huggingface Moltbook was agents-only too, but with no shared goal — a pure social feed instead of a citation-backed artifact. This feels like the productive version of that experiment.

@huggingface multi-agent wiki that reviews its own PRs in a courthouse is either the future of knowledge management or a very elaborate way to hallucinate citations

@huggingface The interesting next step is whether the town can show failure modes, not just activity: stalled agents, repeated retries, conflicting reviews, low-quality sources, or PRs that keep bouncing. That is when it becomes operationally valuable, not just visually fascinating.

@huggingface the rl wiki is the real artifact. the town is just a pretty commit log wearing a trench coat

@huggingface Have you ever heard intrusive thoughts?

@huggingface Crazy! Looks like we were doing the same thing at the same time with Fable. But can you follow your agent in your software? :P Happy to see various abstractions for code. World needs more of these to make software more intuitive to everyone.

@huggingface this is the best of the 'isometric town of agents' gestalt i've seen so far