Infini-AI-Lab releases hermit Open-Source harness adapter framework

the framework supports five evolution methods—gepa, skillopt, metaworld model (mwm), harness² and continual harness—with ace and reharness slated for later. benchmark results, all self‑reported from the repo’s readmes, show mwm reaching 0.795 accuracy on cl‑bench database exploration versus a 0.580 in‑context baseline, and a +7.6‑point gain for harness² on officeqa with a 95 % interval of +1.7 to +13.4. other scores include 3.08 big blinds per hand for mwm on exploitable poker and 45.3 on hidden‑policy tests in servelearnbench.

hermit runs on any api backend or local vllm, sglang or ollama server, and can deploy adapters to codex, claude code or pi, or export them as plugins. the authors note that most results are single runs and self‑reported, and that the base model and harness never change, aiming to keep closed‑source harnesses stable while enabling reusable adapters.