Evidence at a glance
The Adaptation Target Is the Agent's Outer Layer
Infini-AI-Lab released Hermit v0.1.0 on October 25, 2026. It is an open-source harness adaptation framework for AI agents, intended to help agents such as Codex or Claude Code take on new tasks without fine-tuning a model or forking an existing harness. The project is licensed under Apache-2.0 and ships as a Python framework installable with `pip install hermit-harness`.
Hermit's “adapter” is not a model-weight patch like LoRA. It is an editable outer layer around the agent that can contain instructions, skills, memory, tools, executable code, or even an entire repository. The model parameters and base harness remain frozen. That boundary makes changes easier to inspect, but it also means Hermit changes how an agent organizes its work, not the model's underlying knowledge or reasoning ability.
The Feedback Loop Trains Configuration, Not Weights
Hermit organizes adaptation as an RL-style loop. A Trainer runs a batch of tasks through the agent using the current adapter, a judge scores the outputs, and an Evolver reads the trajectories and feedback to propose a new adapter. The Trainer checks that proposal against a template before committing it as the next version. Learning happens in the external configuration, so the quality of the Evolver and its evaluation signal determines whether feedback produces a better adapter.
Each adapter version is stored in a Git tree, making it possible to inspect changes, roll back, or deploy a particular version. Hermit currently includes five evolution methods: GEPA, SkillOpt, MetaWorld Model (MWM), Harness², and Continual Harness. Developers can also write their own Evolver, while ACE and ReHarness are listed as planned additions. Learner programs can run in a sandbox, but that option requires Linux and bubblewrap, which limits how readily it fits into every development environment.
The Results Improve, but the Comparisons Are Not Uniform
All published figures use GLM 5.3 Flash as the acting model and come from repository example READMEs; they are self-reported by the team. On OfficeQA, which scores exact matches across 172 questions about U.S. Treasury Bulletins, the no-adapter baseline scores 0.541, an initial skill 0.576, GEPA 0.580, SkillOpt 0.600, Continual Harness 0.634, and Harness² 0.651. GEPA and SkillOpt train on 50 questions and are then evaluated frozen. Continual Harness and Harness² instead learn while streaming through the test set once, so these scores are not a clean ranking under identical training conditions.
On CL-Bench database exploration, MWM reports 0.795 accuracy against a 0.580 in-context baseline. On exploitable poker, it reports 3.08 big blinds per hand, compared with 2.51 for in-context learning. The results are not uniformly positive: on the same database exploration task, a prompt-only adapter scores 0.470, below the 0.580 baseline. The material also says ServeLearnBench includes 624 retail support requests, but provides no result figures, so it does not support a conclusion about performance on that benchmark.
Portability Still Needs to Be Tested in Real Workflows
Leaving both the model and harness untouched gives teams a relatively low-intrusion way to experiment. The offline Quick Start needs no GPU, model account, or dataset. Real runs can use an API backend or a local vLLM, SGLang, or Ollama server. A trained adapter can be placed in an agent workspace or exported as a native plugin, giving it a possible route into existing deployment workflows without replacing the entire agent stack.
However, portability across compatible models and harnesses is a design rationale, not a broadly demonstrated result in the available material. Version control and performance evidence should also be kept distinct: a Git history shows what changed, but it cannot resolve uncertainty from single runs, different training paths, or self-reported evaluations. For a technical lead, the prudent approach is to treat Hermit as an experimentation and delivery framework for outer-layer policies. Establish a frozen baseline, repeat runs, and use independent review on your own tasks before deciding whether to deploy an adapter in production.