# Hermes Loops

A set of small, reviewable control loops for Hermes Agent installations: one that proposes and tests repository changes in a propose-test-keep-or-revert cycle, and read-only loops that audit a Hermes install's health and suggest improvements for human review.

Hermes Loops packages autonomous agent loops that run against a bounded target and either improve it or leave it unchanged. The autoresearch loop repeatedly lets an agent change a Git repository, measures whether each change improved a score, and keeps good changes as local commits or reverts the rest; it never pushes anywhere and is provider- and agent-agnostic.

The maintenance package ships two independent, read-only loops for a Hermes installation: a health loop that runs bounded probes and reports healthy, warning, unavailable, or unknown findings, and an improvement loop that redacts a supplied outcome and turns repeated patterns into a suggestion for human review, never applying anything automatically.

A third pattern, reviewing a session's own transcript for recurring mistakes, is LLM-driven and ships as a skill in a separate package rather than as code here. Each package installs and versions independently with `pip install`.

Website: https://github.com/AtlasOmnia/hermes-loops