Hermit trains the scaffolding around an AI agent - instructions, skills, memory, tools, code - while leaving the model and the harness itself untouched
AspireAI community · Infini-AI-Lab (GitHub)
Most agent improvement happens by editing the prompt or retraining the model. Hermit takes a third route: keep the model frozen and the existing harness untouched, and train an external adapter around them - a harness adapter that can hold instructions, skills, memory, tools, executable code and agent coordination logic, or be an entire repository implementing an external harness.
The project is an Apache-2.0 framework, released as v0.1.0 on 26 October 2026 and pushed again on 29 October, that provides an RL-style training loop over those adapters plus evaluation, versioning and deployment. It supports five harness-evolution methods today - GEPA, SkillOpt, MetaWorld Model, Harness squared and Continual Harness - with ACE and ReHarness listed as coming soon. A trained adapter converts into a native plugin for Codex, Claude Code, pi or Hermes, and it installs from PyPI as hermit-harness on Python 3.12 or newer.
What this is not is a benchmark result. The repository publishes no task-performance table and no comparison against another harness-evolution method, so the case for harness adapters over prompt edits or model training rests on the design argument rather than measured gains: the original harness can be a closed-source black box, and the adapter stays transferable across compatible models and harnesses. The 0.1.0 tag also carries breaking unreleased changes to the adapter layout - the old 0.1 layout no longer loads - and the README notes that sandboxed agent Evolvers need Linux with bubblewrap.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.