Proteus skill
Agent self-improvement framework: optimizes prompts, memory, skills or harness code with method and evaluation set separately.
- 112 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- Experiments run candidates, including modified harness source, and spend model tokens.
Install
- Works in
Manual install
git clone --branch main https://github.com/proteus-evolve/Proteus.gitFirst command of the quick start. Then create a virtual environment and install the package with `python -m pip install -e .`.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Experiments run candidates, including modified harness source, and spend model tokens.
This is third-party code. Review the repository files before installing.
What it does
A modular framework where the optimization method, agent runtime and evaluation protocol are set independently. The agent proposing changes and the agent executing a candidate can use different harnesses and models: Pi, DeepSeek Harness, Codex, an HTTP adapter or your own integration. Included combinations are Combee с ACE и Combee с GEPA and explicit self-evaluation stages. Evaluation mixes task verifiers, model judges, visible feedback, hidden tests and agent self-assessment. Two skills ship in the repository: `proteus-use` chooses a method, binds roles, runs and resumes experiments, and `proteus-contribute` guides extending the framework. A visual replay of the evolution is included.
Who it is for. Agent researchers and engineers comparing self-improvement methods who want control over which evaluation evidence guides the next update.
Good fit when
- A researcher compares two prompt optimization methods on one task set.
- A team checks whether hidden tests help improve agent harness code.
- An engineer asks the agent to set up an experiment and explain results from the data.
- A developer writes a new stage or evaluator through public contracts.
Not a fit when
- You need a ready improver for one agent without setting up experiments.
- There is no budget for model calls in runs.
- You only need a light tweak to a single prompt.
Example request
Use the proteus-use skill: pick a method to improve my prompts, bind Codex and run the offline comparison, then explain the results.Limitations
Install from `main` with Python 3.10 or newer. Skills ship through Git separately from the Python package: install them from the same tag or commit as the framework. Self-improving harness source runs candidates on your data and spends tokens. The offline comparison example needs no API key or Docker, but real runs depend on the chosen harnesses. Version 1.0 is built from `main`, so check GitHub releases before serious runs. Experiments with hidden-test evaluation need task sets prepared in advance.
How to disable. Remove the `proteus-use` and `proteus-contribute` folders from the agent's skills directory and the virtual environment with the package. Clean experiment result folders manually.
FAQ
How do I install the skill?
In Codex ask the built-in installer for `skills/proteus-use` from `proteus-evolve/Proteus`, ref `main`. In another harness copy the skill folder into its skills directory.
Is an API key needed?
Not for the offline candidate-comparison example. Live experiments need the chosen harnesses and models.
Which harnesses are supported?
Pi, DeepSeek Harness, Codex, an HTTP model adapter or your own integration in supported roles.
What can be optimized?
Prompts, memory, skills and harness source code. The optimization method is chosen separately.
How do the two skills differ?
`proteus-use` runs and explains experiments. `proteus-contribute` guides extending the framework and preparing a contribution, and is optional.
Discussion
No questions or reviews yet.
Related
SkillsA resumable workflow for long-running tasks and a skill platform: from requirements to delivery, with phase checks and skill evaluation
A GEPA plugin for Claude Code: the agent builds an evaluator and searches for a better prompt, code or config by score.
An agent evolution framework that edits prompts and skills with parallel workers and keeps only what passes a held-out check.
A skill that scores a prompt on five criteria, picks one of 31 frameworks and rewrites the request after clarifying questions
Parallel self-evolution of skills and prompts: workers propose edits and an aggregator merges them into one versioned artifact
A Claude Code plugin with hooks and 26 skills: measures what the agent knows, blocks edits until code is read, keeps memory
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| Proteus (this card) | no | 112 | |
| Self-Improving Agent | no | 790 | a self-improvement skill for one agent; Proteus separates method, runtime and evaluation and compares methods |
| Evolving Agents, director | no | 458 | directs agent evolution; Proteus adds ready recipes, evaluators and a visual replay |
README in short
A technical report PDF and an interactive demo of an evolution replay accompany the project. The second skill, `proteus-contribute`, is optional for submitting a PR.