featherloop
v0.1.2
Published
Minimal agent execution loop with configurable, context-lean toolsets
Readme
featherloop
Minimal agent execution loop with configurable, context-lean toolsets.
TBD: screenshot
featherloop is an agent harness designed to bring big possibilities to tiny models: (relatively) big results from small amounts of memory and compute.
Features:
- advisor - let your model ask a higher-tier model for advice
- tool result compaction - summarize some tool results to keep the context short
- extremely small harness size
- compatible with Anthropic and OpenAI APIs; tested with llama.cpp
- under 4k lines of code, 50 kB package size
Motivation
It's the antithesis of tokenmaxxing, which I despise with all my heart, and for which I believe we'll pay a serious environmental price. Beyond that, I've been hosting tiny models on my laptop and wanted to get the most out of them.
With model training moving so incredibly fast, it occurred to me that harnesses carry a lot of instruction models have already learned in training, so I left it out. The success rates so far bear this out.
