🏆 State-of-the-Art Performance on SWE-bench Verified: 80%
- 2025.08.28 🎉 We open-sourced RepoMaster — an AI agent that leverages GitHub repos to solve complex real-world tasks.
- 2025.08.26 🎉 We open-sourced GitTaskBench — a repo-level benchmark & tooling suite for real-world tasks.
- 2025.08.10 🎉 We open-sourced SE-Agent — a self-evolution trajectory framework for multi-step reasoning.
🔗 Ecosystem: RepoMaster · GitTaskBench · SE-Agent · Team Homepage
SE-Agent is a self-evolution framework that enables information exchange between reasoning paths through a trajectory-level evolution mechanism, breaking the cognitive limitations of single trajectories. This approach expands the search space, escapes local optima, and demonstrates emergent capabilities from collective interactions, achieving 80% Top1 performance on SWE-bench Verified, truly realizing autonomous evolution for LLM-based agents in complex reasoning tasks.
SE-Agent ranks Top1 among open-source frameworks on SWE-bench Verified.
SE-Agent gain significant gains on SWE-bench Verified across various open-source/closed-source LLMs.
Get SE-Agent running in 30 seconds:
# 1. Clone and install
git clone https://github.com/JARVIS-Xs/SE-Agent.git
cd SE-Agent
pip install -e .
# 2. Set up API key
echo "DEEPSEEK_API_KEY=your_key_here" > .env
# 3. Run demo (no API calls)
python SE/basic_run.py --mode demo
# 4. Run your first experiment
python SE/basic_run.py --mode executeExpected Output:
✅ SE-Agent initialized successfully
🔄 Starting self-evolution with 3 iterations
💡 Need detailed setup? See 📦 Installation & Configuration below
SE-Agent implements three core self-evolution operations that transform how agents approach problem-solving:
Analyzes individual failed trajectories through deep self-reflection and targeted improvement. Goes beyond simple retries by identifying fundamental approach limitations and creating architecturally orthogonal problem-solving paradigms. This involves analyzing a single trajectory to identify errors, inefficiencies, or conceptual blind spots, then prompting the agent to generate completely different solution approaches that address these specific limitations.
Creates novel trajectories by intelligently combining strengths from multiple existing solution paths. This is where cross-trajectory inspiration primarily occurs - SE-Agent intelligently selects high-performing segments from different trajectories and merges them to construct superior approaches. The process explicitly leverages the interdependence of various attempts, allowing successes in one area to compensate for shortcomings in others, enabling 1+1>2 synergistic effects that transcend individual trajectory limitations.
Optimizes promising trajectories by eliminating redundancies and enhancing efficiency using insights from the entire trajectory pool. After new trajectories are formed, this step further hones them by removing unnecessary steps, streamlining action sequences, and incorporating risk-aware guidance that prevents systematic blind spots and failure modes learned from the collective exploration history.
# Configure multi-iteration strategy
strategy_config = {
"iterations": [
{"base_config": "baseline", "operator": None},
{"base_config": "enhanced", "operator": "alternative_strategy"},
{"base_config": "enhanced", "operator": "crossover"}
]
}Run self-evolution process
python SE/basic_run.py --config SE/configs/se_configs/experiment.yaml --mode executeSE-Agent supports flexible operator extensibility for creating custom evolution strategies:
from SE.operators import TemplateOperator, register_operator
class MyEvolutionOperator(TemplateOperator):
def _generate_content(self, instance_info, problem_description, trajectory_data):
# Implement your custom evolution strategy
return "Your generated strategy content"
# Register and use
register_operator("my_operator", MyEvolutionOperator)📖 Complete Operator Development Guide: See SE/operators.md for comprehensive operator development documentation including architecture, examples, and best practices
# Process multiple SWE-bench instances
sweagent run-batch \
--config config/default.yaml \
--agent.model.name deepseek/deepseek-chat \
--instances.subset verified \
--instances.slice :10SE-Agent provides comprehensive documentation for different use cases:
| Document | Purpose | Audience |
|---|---|---|
| SE/README.md | SE Framework detailed guide | Developers & Researchers |
| SE/operators.md | Operator development guide | Advanced developers |
| instruction.md | Usage instructions & configuration | All users |
- 🚀 Getting Started: Follow the Quick Start above
- ⚙️ Configuration: See instruction.md for detailed setup
- 🔧 Development: Check SE/README.md for framework internals
- 🛠️ Custom Operators: Refer to SE/operators.md for operator development
SE-Agent consists of three main components working in harmony:
📁 SE-Agent Architecture
├── 🧠 SE Framework (SE/)
│ ├── Multi-iteration experiment orchestration
│ ├── Self-evolution operators (Revision, Recombination, Refinement)
│ └── Intelligent trajectory processing & compression
├── 🔧 SWE-Agent Base (sweagent/)
│ ├── LLM agent implementations
│ ├── Environment interaction layer
│ └── Tool execution system
└── 📊 Trajectory System
├── Compressed trajectory storage (.tra files - 80% size reduction)
├── Cross-iteration knowledge accumulation
└── LLM-driven trajectory analysis & summarization
Option 1: Pip Installation (Recommended)
git clone https://github.com/JARVIS-Xs/SE-Agent.git
cd SE-Agent
pip install -e .Option 2: Conda Environment
git clone https://github.com/JARVIS-Xs/SE-Agent.git
cd SE-Agent
conda create -n SE python=3.12
conda activate SE
pip install -e .Verify Installation:
sweagent --help
python SE/test/run_operator_tests.pyChoose one of the following API providers:
# Create .env file
echo "DEEPSEEK_API_KEY=your_deepseek_key" > .env
# OR
echo "OPENAI_API_KEY=your_openai_key" > .env
# OR
echo "ANTHROPIC_API_KEY=your_anthropic_key" > .env📋 Detailed Configuration Guide: See instruction.md for comprehensive configuration options, strategy parameters, and execution workflows
# Run all tests
pytest
# Run SE framework tests
python SE/test/run_operator_tests.py
# Demo mode (no API calls)
python SE/basic_run.py --mode demo
# Code formatting
ruff check .
ruff format .- 🔧 SE Framework Guide: SE/README.md - Comprehensive SE framework documentation with testing, development workflow, and project structure
- ⚙️ Operator Development: SE/operators.md - Complete operator development guide with architecture details and examples
- 📖 Usage Instructions: instruction.md - Detailed usage instructions, configuration options, and execution strategies
If you use SE-Agent in your research, please cite our paper:
@article{se-agent-2025,
title={SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents},
author={Jiaye Lin and Yifu Guo and Yuzhen Han and Sen Hu and Ziyi Ni and Licheng Wang and Mingguang Chen and Daxin Jiang and Binxing Jiao and Chen Hu and Huacan Wang},
journal={arXiv preprint arXiv:2508.02085},
year={2025},
url={https://arxiv.org/abs/2508.02085}
}This project is licensed under the MIT License - see the LICENSE file for details.
We would like to thank the following projects and contributors:
- SWE-Agent - Our foundation framework, developed by Carlos E. Jimenez, John Yang, Kilian Lieret and team
- SWE-bench - For providing the evaluation benchmark and test datasets that enable rigorous assessment of software engineering AI agents
- litellm - For unified LLM API interface support
- Open source community - For contributions to the advancement of software engineering AI agents
- 📧 Email: quantaalpha.ai@gmail.com
- 🐛 Issues: GitHub Issues
- 💬 Discussions: GitHub Discussions
- 🤖 QuantaAlpha: QuantaAlpha is an elite research team dedicated to advancing AI Agent technology, committed to pushing the frontiers of artificial intelligence.
- 🏠 HomePage:https://quantaalpha.github.io
⭐ If SE-Agent helps your research or projects, please give us a star! ⭐
Made with ❤️ by the QuantaAlpha Team Research Team
