Hermit
Hermit#
Train the instructions, skills, roles, and tools used by an existing agent, using feedback from its task attempts. The learned files form a harness adapter: a versioned directory your agent consumes. Model weights stay unchanged. You bring an acting pipeline, tasks, and a judge; the framework manages training, checks, and records.
These docs describe the 0.1.0 code in this checkout.
Your goal |
Start here |
What you will produce |
|---|---|---|
Run a complete training cycle without model access |
A changed adapter, held-out scores, and a host workspace |
|
Run with your model API |
Acting calls, learner edits, and held-out evaluation on included data |
|
Reproduce a benchmark experiment |
A run using the recipe’s data and model setup |
|
Choose a learning method |
A suitable template, Evolver, and split plan |
|
Improve your own agent |
A pipeline, judge, task description, and run config |
|
Use an existing adapter |
A host workspace or a native plugin |
|
Investigate a run |
Evidence for what ran, changed, failed, or resumed |
|
Point a run at another model |
One place binds a model to Fireworks, OpenAI, DeepSeek, or a local vLLM or SGLang server |
The training loop#
The pipeline runs your acting agent with the current adapter on a batch of tasks.
The judge scores its outputs. The Trainer collects scores and available trajectories as learning material.
The Evolver uses this evidence to revise the adapter’s learnable files.
The Trainer checks the candidate, records the accepted version, and starts the next batch.
You evaluate a selected adapter on held-out tasks and deploy it to an agent.
Passing file checks establishes structural validity. Performance improvement requires evaluation. The concept guide defines the package, and the interactive overview shows its relationships.
Get Started
User Guide
Learning Methods
Recipes and Reproduction
Developer Guide