Overview

Hermit improves the instructions, skills, and tools used by an existing AI agent through feedback from task executions. It packages those additions as a harness adapter, and provides the loop for running tasks, learning from results, and saving revised adapters.

Using an adapter

An adapter provides context, skills, roles, and callable tools to an existing agent. An optional independent service can exchange information with the agent and coordinate its execution through a host integration. The available interactions depend on the service and host capabilities. Click a component below to see its role and connections.

Harness adapterA package of capabilities

Acting agent(s) / hostOne or more agents with an existing harness

ContextInformation and instructions Skills + rolesGuidance and role definitions ToolsCallable capabilities Independent serviceBackground logic + coordination Acting agent(s)Existing harness + model(s)

Tool calls are initiated by the agent. An independent service can also exchange information and coordinate execution through the host integration.

An instruction-only adapter uses just the context path. Using a saved adapter does not require the Trainer or Evolver.

Evolving an adapter

Hermit repeats a feedback loop: run the acting agent with an adapter, collect learning material, let an Evolver revise a working copy, then check and save the version used by the next batch.

Training/Evolution LoopRun · learn · validate

Acting agent(s)

Task data Original harness Harness adapter Specification check Learning material Environment EvolverEvolver state Validation Human
feedback

Solid: main loop. Dashed: optional access and inputs.

The task owner supplies data, the acting pipeline, and any scoring or feedback sources. The template defines which adapter files may change. Human feedback can be supplied through task material or tools; it is not a separate built-in feedback service. Passing the specification check establishes validity; improved task performance must be measured separately.

Scoring feedback is included on the Agent → Learning Material path. The current train and eval implementation still requires a scoring callback through judge.path.

Start with Installation and Quick Start. For file layouts and exact handoffs, see Using and training adapters. Component details: Adapter, Evolver, Trainer, and Task.