We are the Frontier
World

New Hermit Framework Enables Modular AI Agent Training

The open-source tool allows developers to train and track agent adapters without modifying the underlying model harness.

Updated on Oct. 29, 2026 in Artificial Intelligence

Isometric editorial illustration showing a modular geometric component connecting to a central block, representing software framework architecture.
Infini-AI-Lab released its Hermit framework, an open-source tool for training and porting AI agent adapters without altering primary model source code. AI Illustration. Upload story photo >

Live Poll

Would you prefer to use external plugins to adapt your AI agents to new tasks?

Infini-AI-Lab has released Hermit v0.1.0, an open-source Python framework designed to train external adapters for AI agents without modifying the underlying model or harness. This initial release enables users to improve agent performance through five distinct evolution methods while maintaining auditability via git-based version control.

Why it matters

By decoupling adapters from proprietary agent harnesses, Hermit addresses the fragility of current agent architectures where modifying source code often breaks component interactions. This modularity allows researchers to retrain and port learned behaviors across different agent models.

Using the GLM 5.3 Flash model, the MetaWorld Model (MWM) adapter achieved an accuracy of 0.795 on database exploration tasks, outperforming the 0.580 baseline. In poker simulations, the MWM adapter secured 3.08 big blinds per hand compared to the 2.51 baseline.

The players

Infini-AI-Lab

A research entity focused on the development of open-source frameworks for agent-based machine learning.

Haizhong Zheng

A researcher who announced the release of the Hermit framework.

The details

Hermit functions by freezing the primary model and harness, then training an external adapter using a feedback loop. This system runs tasks through the agent, has a judge score the output, and employs an evolver to propose improved adapter versions. These versions are tracked within a git tree, a file management system used for tracking changes in code, allowing developers to inspect or rollback adapter states during the training process.

Timeline

  1. October 25, 2026: Hermit v0.1.0 was released.

  2. October 27, 2026: Haizhong Zheng announced the framework on X.

The Tech Race

Hermit seeks to standardize the adaptation of AI agents by providing an open-source alternative to opaque, closed-source harnesses. It advances the field beyond fixed in-context baselines by enabling iterative, version-controlled adapter evolution.

Developers and researchers can immediately integrate Hermit v0.1.0 with compatible agents like Codex, Claude Code, and Pi. The framework requires familiarity with Python and Git-based versioning to manage adapter training and rollback workflows.

The takeaway

Hermit provides a structured way to improve agent performance without the risks associated with modifying proprietary harnesses. Researchers should monitor the Infini-AI-Lab repository for the future integration of the ACE and ReHarness evolution methods.

Further reading

Learn more about the latest innovations in Artificial Intelligence.

Live Poll

Would you prefer to use external plugins to adapt your AI agents to new tasks?