Infini-AI-Lab Releases Open-Source AI Agent Harness Adapter

The first version, v0.1.0, was released on October 25, 2026. CMU postdoc Haizhong Zheng announced the release on X on October 27, 2026. Hermit aims to enable developers to adapt agents like Codex or Claude Code to new tasks without requiring fine-tuning or forking the harness.

Hermit is a Python framework with no parameters of its own and is available under an Apache-2.0 license, installable via pip. The offline Quick Start version requires no GPU, model account, or dataset. For actual use, it can run with any API backend or a local vLLM, SGLang, or Ollama server.

Performance reports indicate gains over baselines without adapters on benchmarks like OfficeQA, CL-Bench, and ServeLearnBench, using GLM 5.3 Flash. The MWM adapter achieved 0.795 accuracy on CL-Bench database exploration, surpassing the in-context baseline of 0.580. However, the prompt-only adapter scored 0.470 on the same benchmark.

A harness adapter functions as a trainable external component, similar to a plugin with integrated training support. Unlike model adapters like LoRA that adapt a model, a harness adapter adapts an existing agent. It can incorporate instructions, skills, memory, tools, executable code, and agent coordination logic.

The design choice to keep the base model and harness static is attributed to several factors: many harnesses are closed-source, editing open-source ones can disrupt fine-tuned component interactions, and a learned adapter can be transferred between compatible models and harnesses. Hermit employs an RL-style training loop for adapter training.