Without Modifying the Base Model or Harness, Hermit Trains an Agent External Adapter Layer Using Task Feedback
Infini-AI-Lab open-sources Hermit under the Apache-2.0 license. It trains a dedicated external capability layer covering instructions, skills, memory and tools from task feedback, while leaving the original model and harness completely untouched. For Agent engineering, this delivers an iteration path that requires minimal modifications to the underlying base system.
Infini-AI-Lab has released Hermit, a harness adapter framework licensed under Apache 2.0. Asif Razzaq from MarkTechPost covered this launch. The core value proposition can be summed up in one line: train a separate external capability layer that includes instructions, skills, memory and tool usage directly from task feedback, without modifying either the Agent's underlying model or its existing harness.
The harness can be understood as the scaffolding of an Agent: it is responsible for organizing prompts, tool calling, memory read/write operations and control loops. Instead of replacing this existing scaffolding, Hermit adds an extra trainable adapter layer on top of it.
The highlight of this design is that when an Agent delivers subpar performance, the common industry practices include adjusting prompts, switching to a different model, rewriting tool orchestration logic, or performing fine-tuning. Hermit pushes all modifications to the outer layer, keeping the core model and harness fully stable, while the outer layer can be iterated independently using task feedback. For engineering teams, this makes rollbacks and component replacement far easier, at least per the framework's intended design.
The Apache 2.0 license also lowers the adoption barrier: commercial use is permitted, as long as users retain the required license notices and terms as specified. More critically, the quality of training signals — how task feedback is collected, how performance quality is evaluated, and whether the trained outer layer can generalize to new tasks — will determine how practical Hermit is in real-world use. Public reports currently only clarify the framework's positioning, and no benchmark results or full experimental data have been released yet.
If you are building Agent products, you can consider this as a viable architectural option that leaves the core base system unchanged while supporting trainable outer layers. Do not treat it as a universal quick fix right away; start by reviewing its code, interfaces and the associated implementation and reproduction costs.
Original link: Infini-AI-Lab Releases Hermit, an Open-Source Harness Adapter Framework for AI Agents
发布时间: 2026-10-29 16:24