CMU's Hermit: Unified Framework for Fair Harness Evolution Comparison

Problem: Unfair Comparison of RSI Methods

Research on Recursive Self-Improvement (RSI) has surged, but different methods typically use their own benchmarks, training configurations, and evaluators. This makes fair comparison difficult — akin to athletes setting their own rules and acting as referees.

Key Insight: Shared Training Logic

CMU researchers observed that despite implementation differences, various Harness evolution methods follow a similar training loop: let the Harness execute tasks, collect feedback, have the RSI algorithm propose modifications based on feedback, validate the changes, save a new version, and iterate.

Hermit: A Unified Training Framework

To enable fair comparison under consistent conditions, CMU open-sourced Hermit — a unified framework for Harness evolution, continuous learning, and RSI. It externalizes the learning process via pluggable Harness Adapters that continuously improve the agent without modifying the underlying model or the original Harness.

Why Harness Adapters?

Many Agent Harnesses are black boxes with carefully designed internal component interactions. Directly modifying internal logic risks breaking existing workflows. Hermit instead uses supported extension interfaces to introduce trainable Adapters externally. Even if both the model and Harness are black boxes, as long as they provide extension interfaces, Harness evolution becomes possible.

Three Advantages of the Adapter Design

No modification of existing Harness — lowers retrofitting cost and reduces risk from changing internal components.

Adapter portability — adapters can be reused across compatible models and Harnesses, avoiding starting from scratch when switching systems.

Continuous training via task feedback — adapters turn execution experience into external configuration improvements.

Complete Feedback Loop

Hermit implements a closed loop: the Agent executes a task, an Evaluator checks the result, an Evolution Algorithm proposes modifications based on feedback, the changes are validated, saved as a new Adapter version, and the next iteration begins.

Each accepted update becomes the next adapter version.

Supported Evolution Algorithms

Hermit does not bind to a single optimization method. It already supports multiple algorithms including GEPA , SkillOpt , MetaWorld Model , Harness² , and Continual Harness . Researchers can plug in different algorithms under a relatively unified training pipeline, keeping the model, tools, evaluator, and benchmark consistent for fairer empirical comparison.

Tutorial: First Training Example

The following steps require Python 3.12 and Git. Clone the Hermit repository and set up the environment:

cd hermit
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .

Run the first training example (no GPU, no model API account, no dataset download needed):

# 1. Check configuration
hermit check --config docs/_examples/first_training/run.yaml

# 2. Evaluate before training
hermit eval --config docs/_examples/first_training/run.yaml --adapter none

# 3. Start training
hermit train --config docs/_examples/first_training/run.yaml

# 4. Evaluate after training
hermit eval --config docs/_examples/first_training/run.yaml --adapter final

The task is simple: the Agent must output uppercase text. Initially the Agent only outputs lowercase, failing evaluation (score 0.0). During training, the learning module reads evaluation feedback and writes improvement rules into ADAPTER.md, instructing the Agent to use uppercase format. After training, evaluation score rises to 1.0.

Inspect the learned rules:

cat .demo/first-training/first-training/checkout/ADAPTER.md

Deploying Learned Rules to Codex

To deploy the learned adapter to a Codex workspace:

hermit deploy \
  --adapter .demo/first-training/first-training/checkout \
  --workspace .demo/first-training/workspace \
  --host codex

This generates an AGENTS.md file in the workspace. Codex will then follow the learned rule for future tasks.

The complete Hermit flow — task execution, feedback collection, adapter update, re-evaluation, and deployment — is now verified end-to-end.