Infini-AI-Lab has released Hermit, an open-source framework that trains an external “harness adapter” for AI agents — a way to adapt agents like Codex or Claude Code to new tasks without fine-tuning the model or forking the harness. The first release, v0.1.0, shipped on October 25, 2026, and was announced on X by CMU postdoc Haizhong Zheng on October 27.
Hermit (Credits: MarkTechPost)
What Is a Harness Adapter?
A harness adapter is a trainable external harness — think of it as a plugin with training support built in. Where a model adapter like LoRA adapts the model itself, a harness adapter adapts an existing agent instead. The adapter can hold instructions, skills, memory, tools, executable code, and even agent coordination logic. Its specification defines which parts are learnable and how updates get checked — while the base model and harness never change.
The team gives three reasons for this design: many harnesses are closed-source, editing open ones can break carefully tuned component interactions, and a learned adapter can move between compatible models and harnesses.
How the Training Loop Works
Hermit uses an RL-style training loop. A Trainer runs a batch of tasks through the acting agent with the current adapter, a judge scores the outputs, and an Evolver reads the trajectories and feedback to propose a new adapter — which the Trainer checks against the template and commits as the next version. Every adapter version lives in a git tree, so you can inspect, roll back, or ship any version, and learner programs can run sandboxed on Linux with bubblewrap.
Five evolution methods ship today: GEPA, SkillOpt, MetaWorld Model (MWM), Harness², and Continual Harness — with ACE and ReHarness listed as coming soon, plus support for custom Evolvers.
The Numbers Behind Hermit
Hermit is a pure Python framework with no parameters of its own, licensed Apache-2.0 and installable via pip install hermit-harness. The offline Quick Start needs no GPU, model account, or dataset; real runs use any API backend or a local vLLM, SGLang, or Ollama server.
The reported results are promising but come with a caveat — most are single runs and self-reported. On CL-Bench database exploration, the MWM adapter hit 0.795 accuracy versus 0.580 for the in-context baseline. On OfficeQA, Harness² reached 0.651 against a 0.541 no-adapter baseline. On ServeLearnBench, MWM scored 45.3 on hidden-policy tests — well ahead of the leaderboard’s best learning method at 28.7, though still far from the oracle’s 89.2.
Also Read: NVIDIA Releases Nemotron 3.5 Transcribe: Open 30B MoE Model for 40-Language Transcription
Also Read: Adam Mosseri Says AI Negativity Is Going to Get Worse, Driven by Fear of Being Left Behind
Source: MarkTechPost — read the full story there.
