Hermit adapts AI agents through an external layer

Infini-AI-Lab’s Hermit offers developers a way to adapt agents without changing their underlying model or harness, the infrastructure that runs the agent. Instead, feedback shapes an external adapter: a trainable extension containing instructions, skills, memory, tools or coordination logic.

The real-model tutorial separates training from checking the result: nine sentiment-classification tasks are split equally into training, validation and held-out testing. It keeps the acting model’s settings fixed when comparing the baseline, initial adapter and final adapter. Readers should check that execution errors are zero; a low score without such errors can still provide valid training feedback. Three test tasks cannot establish a reliable improvement, and a learned change can make results worse.

The deployment guide draws another boundary: generating workspace files does not start the host application. The workspace must be empty or previously generated for that same host, and separate from the adapter; an arbitrary nonempty project directory is refused. Redeployment replaces generated configuration and the managed context block, so those generated parts should not be treated as preserved manual edits. Converting an adapter into a plugin package also does not install or enable it, and file generation alone is not a performance measurement.

MarkTechPost provides a secondary account of the project and its external-adapter approach.