You guys, you seriously need to check out AmazingDigitalPetDentures — it's probably my favorite hackathon project I've seen in ages! It started as a proof-of-concept and now they just hit v1.0 on Hugging Face with their own demo page at hf.co/AmzingDigitalPetDentures where you can actually see the world cards expand in real time (84 of them generated already). The core idea is wild: it's a digital pet (in the form of... a denture, yeah) that feeds off your daily "Adventures" and evolves as you feed it. Each day you input an adventure — whatever happened to you — and instead of just logging text, the system routes that through a graph of specialized AI agents. There's a researcher agent to analyze the input, a writer who drafts new content, and even critic agents that review every step before anything goes live. This isn't some single massive LLM call; it's multiple smaller steps intelligently choreographed so each one can be tiny and cheap while still producing high-quality output together. That kind of modular thinking is exactly where the future of agentic workflows lives, in my opinion — building a pipeline of specialists rather than expecting one model to do everything at once.
Under the hood it's built on Zero Agents 1, which I can tell you is genuinely clever because it solves one of the hardest problems in LLM app design: planning and decomposition without going broke on API fees. The orchestrator deconstructs your prompt into a directed graph where each node is an agent with its own system instructions, and they communicate through structured handoffs — no ambiguous chat history mess between steps. This lets them use small, fast models like Phi-2 or Gemma 7B instead of dropping GPT-4 tokens on every microtask. They even have error recovery built in, so if one step fails it doesn't crash the whole flow; the orchestrator just reroutes and retries a different path. The output isn't just text either — it generates new cards with backstory, lore expansions, and world connections that get added to your personal collection every time you add an adventure. It transforms simple journaling into dynamic world-building that grows exponentially over time because each old card can seed new ones through the graph! That kind of emergent narrative is exactly why I'm obsessed with this architecture — it takes a tiny input and multiplies it through smart orchestration rather than just regurgitating it.
For those interested in the broader picture, this project was part of Hugging Face's "build small" hackathon series alongside SmolChat23 and their SmallModelShowcase -- all focused on getting state-of-the-art results from tiny models instead of bigger ones. You can check out the original blog post at hf.co/blog/build-small-hackathon for more context, plus there are several related projects like a local LLM search engine and an audio summarization tool that emerged from the same challenge. But Seriously, everyone go to hf.co/AmzingDigitalPetDentures and try it yourself — start your first adventure and watch how the system expands on it in seconds! The whole project is open source too so you can dive into the code if you want to see exactly how Zero Agents handles the routing under the hood. I'm already planning my own agent-based system inspired by this approach, because honestly once you see what a well-designed orchestration graph can do with small models it's impossible not to build your own thing!
Source: https://huggingface.co/blog/build-small-hackathon/amazingdigitaldentures
Under the hood it's built on Zero Agents 1, which I can tell you is genuinely clever because it solves one of the hardest problems in LLM app design: planning and decomposition without going broke on API fees. The orchestrator deconstructs your prompt into a directed graph where each node is an agent with its own system instructions, and they communicate through structured handoffs — no ambiguous chat history mess between steps. This lets them use small, fast models like Phi-2 or Gemma 7B instead of dropping GPT-4 tokens on every microtask. They even have error recovery built in, so if one step fails it doesn't crash the whole flow; the orchestrator just reroutes and retries a different path. The output isn't just text either — it generates new cards with backstory, lore expansions, and world connections that get added to your personal collection every time you add an adventure. It transforms simple journaling into dynamic world-building that grows exponentially over time because each old card can seed new ones through the graph! That kind of emergent narrative is exactly why I'm obsessed with this architecture — it takes a tiny input and multiplies it through smart orchestration rather than just regurgitating it.
For those interested in the broader picture, this project was part of Hugging Face's "build small" hackathon series alongside SmolChat23 and their SmallModelShowcase -- all focused on getting state-of-the-art results from tiny models instead of bigger ones. You can check out the original blog post at hf.co/blog/build-small-hackathon for more context, plus there are several related projects like a local LLM search engine and an audio summarization tool that emerged from the same challenge. But Seriously, everyone go to hf.co/AmzingDigitalPetDentures and try it yourself — start your first adventure and watch how the system expands on it in seconds! The whole project is open source too so you can dive into the code if you want to see exactly how Zero Agents handles the routing under the hood. I'm already planning my own agent-based system inspired by this approach, because honestly once you see what a well-designed orchestration graph can do with small models it's impossible not to build your own thing!
Source: https://huggingface.co/blog/build-small-hackathon/amazingdigitaldentures