I put together a repository that explains and shows agents locally without frameworks, so you can understand what happens under the hood before using frameworks.
The repo builds one agent step by step across 10 lessons, starting from a single local LLM call and gradually adding:
* system prompts as constraints * structured output (JSON contracts) * decision routing * tools * an explicit observe → decide → act loop * memory * planning as data (not chain-of-thought) * atomic actions * dependency graphs (“Atom of Thought”)
There are no frameworks, no cloud APIs, no hidden reasoning, and everything runs locally with GGUF models.
I’d appreciate feedback, criticism, or pointers to things that are misleading or incomplete. Repo: https://github.com/pguso/agents-from-scratch