Smol Depot

Smol Depot comes out of the smol.ai ecosystem — the “smol” philosophy that made its name with smol developer, the tiny “junior developer” agent that turns a plain-English spec into a working codebase. The idea across the smol family is the same: agent scaffolds small enough to read in one sitting, meant to be copied into your project and modified, rather than imported as a black-box framework.

That makes this the counterweight to everything else in this module. LangGraph, CrewAI, and AutoGen give you abstractions to learn; the smol approach gives you a few hundred lines of prompt-plus-loop code to own. It’s the practical descendant of the manual implementation lesson: instead of writing your agent from a blank file, you start from a minimal, working template — a depot of small agent starting points — and evolve it. The same minimalist current runs through Hugging Face’s smolagents library, which keeps the entire agent loop in a compact, hackable core. When your task is well-scoped (scaffold an app, transform files, run a research loop), this style often ships faster and debugs easier than a framework.

In practice, you’ll clone or copy a template rather than pip install a heavy dependency, drop in your model API key, and read the whole thing before running it — that’s the point. You’ll edit the system prompt directly, adjust the loop’s tools and stop conditions, and treat prompts as source code under version control. Keep what works, delete the rest.