Agent Forge Guild

Build a real AI agent with other people, one small piece at a time

Building a real AI agent usually means one person has to understand the whole thing at once: how a model reads data, how it learns from being wrong, and how it decides what to say back. Agent Forge Guild breaks that whole job into small, named pieces — a forward pass, an activation function like ReLU, a way of picking the next word — so anyone can pick one small piece, learn it well, and see exactly how it fits into the rest.

The standard list below never changes. If you think you have something genuinely better — the way "Attention Is All You Need" once proposed a whole new way to restructure training and inference — post it as a new Idea instead: name it, list which standard section or sections it replaces, and bring its real code with it.

Browsing the catalog and every golden example below is free for everyone, whether or not you have the $16.99/month Easy Life Agent Subscription.

How to use this page — step by step

1. Look through the three groups below: Training Zoo, Inference Zoo, and Agent Zoo.
2. Pick one small piece that sounds interesting, like ReLU or Sampling Strategies.
3. Open its golden example to read one small, working piece of code with a plain explanation next to it.
4. Click a part's card to submit your own piece for it, vote on others', and join its discussion.
5. Have a genuinely new idea instead of a standard part? Post it as an Idea with its own real code and the standard section(s) it replaces, then talk it over.
6. Some parts include a real release's own real code — GPT-2, Llama 2, Mistral 7B. Use the Composer to mix and match sections from different, competing models on purpose — GPT-2's attention next to Llama 2's activation function — for real side-by-side comparison and testing, not because any real model actually runs that way.