DarwinAgent

Evolution for the Agent Era · Agent 时代的进化

An open framework for experience-driven recursive self-improvement of AI agents. DarwinAgent takes its name from Charles Darwin and the theory of evolution: propose changes to task assets, select through independent evaluation, retain effective versions, and use a persistent experience Wiki to inform subsequent proposals.

Version 0.1.0 is experimental. Improvement is an outcome to measure, and a candidate can be rejected. The framework does not train model weights or rewrite its own optimizer.

Install and try the loop

Python 3.11 or newer is required. In your Python environment:

python -m pip install darwinagent
darwinagent --version
darwinagent doctor --output runs/doctor
darwinagent demo --mode replay --rounds 2 --output runs/demo-replay
darwinagent demo --mode replay --rounds 2 --output runs/demo-replay --resume

Replay requires no API key or network. It executes the actual controller, task runtime, evaluation, admission, Wiki, and asset publication with scripted model responses. In one tiny synthetic maintenance task, the first candidate is accepted and the second is rejected on a tie; HTTP attempts are zero. This demonstrates the mechanism and does not establish model learning or benchmark gains.

For real model execution, set DARWINAGENT_API_KEY, DARWINAGENT_BASE_URL, and DARWINAGENT_MODEL to your OpenAI-compatible Chat Completions endpoint, then run:

darwinagent demo --mode live --rounds 2 --output runs/demo-live \
  --max-requests 40 --timeout 1800

Live mode calls the configured model. The request cap counts dispatched HTTP attempts, including retries, and the timeout covers the whole run.

Bring your own task

Register S/F/C/P assets: schemas, query functions, task checks, and role prompts. Connect a dataset adapter and an independent evaluator to the shared Pipeline and ExperimentRunner. The framework's executor, evaluator, permissions, and adoption policy remain outside the proposal boundary.

Documentation

• English README
• 中文说明
• Quickstart
• 中文快速开始
• Acceptance evidence
• Source and issues

DarwinAgent is distributed under the MIT license.