by JARVIS-Xs
SE-Agent is a self-evolution framework for LLM Code agents. It enables trajectory-level evolution to exchange information across reasoning paths via Revision, Recombination, and Refinement, expanding the search space and escaping local optima. On SWE-bench Verified, it achieves SOTA performance
# Add to your Claude Code skills
git clone https://github.com/JARVIS-Xs/SE-AgentNo comments yet. Be the first to share your thoughts!
🔗 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 execute
Expected 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 execute
SE-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 :10
SE-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 |
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.py
Choose 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 .