We propose the EntropyEvolve.
This architecture is a general framework for agents self-improvement, applicable to any complex domain such as science or medicine. However, we have chosen software engineering as our initial testing ground.
We have a base coding agent that we want to improve. Using a continuous improvement cycle, EntropyEvolve is able to continually improve itself with feedback from its errors.
git clone https://github.com/luisjosuecortes/EntropyEvolve.git
git clone https://github.com/SWE-bench/SWE-bench.git
cd SWE-bench
pip install -e .
cd ..
pip install langgraph
pip install openai
python cycle_graph.py
This project develops a self-improving agent system designed to optimize its performance in solving programming problems.
- The system consists of three coding agents, each assigned tasks from the SWE-Bench benchmark.
- An evaluation node executes their solutions, collects the results, and generates logs detailing any detected errors.
- These logs are analyzed by an evaluation agent, which extracts key insights about performance and mistakes.
- Based on this analysis, an optimization agent adjusts the prompts of the coding agents, thereby restarting the continuous improvement cycle.
- The entire system is implemented using a LangGraph graph structure, while the agents themselves are powered by OpenAI’s large language models (LLMs).
We used SWE-bench for testing and having a quantitative evaluation.