While working on an educational exercise tinkering with local models and trying my hand at setting up agents, I went down a rabbit hole: to see how far I could build a custom agent loop using exclusively command-line building blocks and stripping out dependencies wherever possible.

It turns out you can get pretty far with pipes, text streams, append only logs, and standard command-line components - concepts pretty well aligned with classic Unix philosophy. The agent is a wrapper composed of a handful of smaller programs, which should allow for flexibly injecting various tool to inspect, filter, redirect, and audit different stages of the agent loop.

This project as it stands is a proof-of-concept, but packs enough punch with tool calling support - which in theory should make it indefinitely extensible. With that said, it does appear that the sophistication of tool calling is largely limited by the underlying model, and so far I’ve only experimented with lightweight local models (e.g. llama3.2:1b) that have modest success for tools.

Nevertheless, I wanted to share this here and am curious to see if others find it interesting enough to build upon or extend!