Editorial illustration for Infini-AI-Lab Releases Hermit, Open-Source Agent Harness Adapter Framework
Hermit: New Open-Source Agent Framework Released
Infini-AI-Lab Releases Hermit, Open-Source Agent Harness Adapter Framework
Infini-AI-Lab pushed version 0.1.0 of Hermit on October 25, 2026, and two days later CMU postdoc Haizhong Zheng announced it on X. The package is a Python framework, installable with pip install hermit-harness, licensed under Apache-2.0, and it carries no parameters of its own. That last detail matters: Hermit trains an external "harness adapter" while leaving both the agent's underlying model and its harness untouched.
The pitch targets developers who want to point Codex or Claude Code at a new task without fine-tuning the model or forking the harness code. A Quick Start runs offline, with no GPU, model account or dataset required, and real deployments can hook into any API backend or a local vLLM, SGLang or Ollama server. Infini-AI-Lab reports gains over no-adapter baselines across three benchmarks, OfficeQA, CL-Bench and ServeLearnBench, all tested with GLM 5.3 Flash, though most of those numbers come from single, self-reported runs rather than repeated trials.
The core idea hinges on a term the team is introducing alongside the release.
Hermit is a framework from Infini-AI-Lab that trains an external ‘harness adapter’ while the agent’s model and harness stay frozen. The first release, v0.1.0, shipped on October 25, 2026.
Why this matters
For teams running Codex or Claude Code in production, Hermit's pitch cuts against the usual playbook of forking a harness or fine-tuning a model every time a task shifts. Keeping the agent and harness frozen while training a separate adapter means you can version and audit the thing that actually changes, without re-certifying the whole stack each time. That's a real operational win if you're maintaining agents across multiple clients or tasks, since forked harnesses tend to rot the moment upstream ships an update.
We'd want to see how Hermit's adapter spec handles permissions in practice, since "which parts are learnable" is doing a lot of work in that sentence. Still, v0.1.0 landing October 25 with Haizhong Zheng pushing it publicly two days later suggests CMU's lab wants outside eyes on this fast. Worth watching whether other agent frameworks adopt the same frozen-core pattern, or whether this stays a niche tool for teams already deep in the Codex/Claude Code ecosystem.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives GitHub Copilot Projects Adds Task Tracking to Code Assistant
From Archives Nvidia releases new open transcription model to recognize industry jargon
Common Questions Answered
What is the key advantage of Hermit's harness adapter approach compared to traditional fine-tuning?
Hermit trains an external harness adapter while keeping both the agent's underlying model and its harness frozen, allowing teams to version and audit only the adapter that changes without re-certifying the entire stack. This approach eliminates the need to fork harnesses or fine-tune models every time a task shifts, providing significant operational efficiency for teams managing agents across multiple clients or tasks.
How can developers install Hermit and what license does it use?
Hermit is a Python framework that can be installed using the command 'pip install hermit-harness' and is licensed under Apache-2.0, making it freely available for open-source and commercial use. Version 0.1.0 was released by Infini-AI-Lab on October 25, 2026, and announced publicly by CMU postdoc Haizhong Zheng two days later.
Why does it matter that Hermit carries no parameters of its own?
Because Hermit carries no parameters of its own, it functions purely as a training framework for the external harness adapter without adding computational overhead or model complexity. This design allows the underlying agent model and harness to remain completely untouched and frozen, simplifying deployment and maintenance across production environments.
Which AI models does Hermit support for agent development?
Hermit is designed to work with models like Codex and Claude Code, enabling developers to point these agents at new tasks without modifying the core model or harness infrastructure. The framework's adapter-based approach makes it compatible with various agent architectures that teams may already be running in production.
Further Reading
- Infini-AI-Lab Releases Hermit, an Open-Source Harness Adapter Framework for AI Agents - MarkTechPost
- GitHub - Infini-AI-Lab/hermit - GitHub
- What Is an Agent Harness? Architecture and Setup in 2026 - DigitalOcean
- Best Agent Harnesses 2026: A Builder's Field Guide - Future AGI
- The best AI agent frameworks in 2026 - LangChain
