Infini-AI-Lab Releases Hermit: An Open-Source Harness Adapter Framework for AI Agents
Infini-AI-Lab has released Hermit, an open-source harness adapter framework for AI agents. Hermit trains an external "harness adapter" while the agent's underlying model and harness remain frozen. The first release, v0.1.0, shipped on October 25, 2026, and CMU postdoc Haizhong Zheng announced it on X on October 27. For developers, the pitch is straightforward: adapt Codex or Claude Code agents to new tasks without fine-tuning the model or forking the harness.
TL;DR
- Footprint: A Python framework with no parameters of its own. Licensed under Apache-2.0; install with
pip install hermit-harness. - Runs on: The offline Quick Start requires no GPU, model account, or dataset. Real runs work with any API backend or a local vLLM, SGLang, or Ollama server.
- Performance: Reported gains over no-adapter baselines on OfficeQA, CL-Bench, and ServeLearnBench, all using GLM 5.3 Flash.
- Best result: The MWM adapter reaches 0.795 accuracy on CL-Bench database exploration, versus 0.580 for the in-context baseline.
- Worst result: The prompt-only adapter scores 0.470 on the same benchmark, below the 0.580 baseline.
- Bottom line: Best: a single training loop covering five harness evolution methods. Worst: most results are single runs and self-reported.
What Is a Harness Adapter?
A harness adapter is a trainable extension layer that sits around an existing agent harness. Rather than modifying the model weights or rewriting the harness itself, Hermit learns a lightweight adapter that shapes how the agent interacts with its tools, environment, and task structure. This keeps the base agent intact while allowing it to specialize for new workloads.
The approach is particularly relevant in 2026, as agentic systems built on Codex and Claude Code have become standard tooling for software and research workflows. Fine-tuning large frontier models remains costly and often impractical, and forking a harness introduces maintenance burden. Hermit's frozen-model, trainable-adapter design targets exactly that gap: teams can adapt an existing agent to a new domain without touching the model or the harness code.
Benchmarks and Results
Hermit's reported evaluations span three benchmarks โ OfficeQA, CL-Bench, and ServeLearnBench โ all run with GLM 5.3 Flash. The strongest result is the MWM adapter's 0.795 accuracy on CL-Bench database exploration, compared with 0.580 for the in-context baseline. The weakest is the prompt-only adapter at 0.470 on the same benchmark, which underperforms the baseline.
| Method | Benchmark | Accuracy | Baseline |
| --- | --- | --- | --- |
| MWM adapter | CL-Bench database exploration | 0.795 | 0.580 (in-context) |
| Prompt-only adapter | CL-Bench database exploration | 0.470 | 0.580 (in-context) |
Caveats
Two limitations are worth noting. First, most reported results come from single runs and are self-reported by the authors, so independent replication would strengthen the claims. Second, the prompt-only adapter's underperformance suggests that not every adapter strategy delivers gains; method choice matters.
Getting Started
Hermit is available as a Python package under the Apache-2.0 license. Developers can install it with pip install hermit-harness and run the offline Quick Start without a GPU, model account, or dataset. For real workloads, Hermit supports any API backend as well as local inference through vLLM, SGLang, or Ollama.
With its single training loop covering five harness evolution methods, Hermit offers a practical entry point for teams looking to specialize Codex or Claude Code agents without fine-tuning or forking the harness.
via MarkTechPost