FAST-Evolve:基于编码智能体生态的算法优化进化工具链项目

FAST-Evolve 是面向算法优化的一站式智能体进化工具链。给出算法问题描述和评测标准,它就能驱动 Coding Agent 自主探索算法方案设计,依据评测反馈持续迭代,让更优解不断涌现。

Branch1Tags2
FAST-Evolve

English | 中文

🌱 Overview

FAST-Evolve is an all-in-one agentic evolution toolchain for algorithm optimization. Given a problem description and evaluation criteria, it drives coding agents to autonomously explore and evolve algorithm designs using evaluation feedback, yielding a steady stream of better solutions.

FAST-Evolve currently provides program evolution capabilities for coding agents through EvoTrias. By integrating an optimization task and an evaluator into a workspace, it drives agents to continuously generate, evaluate, and select candidate programs. EvoTrias supports Claude Code and OpenCode. See the EvoTrias documentation for detailed usage.

🚀 Quick Start

This example uses the built-in Circle Packing task, which aims to place 26 non-overlapping circles inside a unit square while maximizing the sum of their radii.

Prepare Python 3.11 or later, and install and configure your chosen coding agent (Claude Code or OpenCode) and its model service. We recommend installing dependencies in a Python virtual environment.

1. Get the Code and Install

git clone https://github.com/openJiuwen-ai/FAST-Evolve.git
pip install ./FAST-Evolve/evo_trias
pip install -r FAST-Evolve/benchmarks/requirements/math.txt
evotrias --version

2. Create a Workspace and Import the Task

Create a workspace alongside the repository and initialize EvoTrias:

mkdir my-workspace && cd my-workspace
evotrias setup

Follow the prompts to select a coding agent and complete the installation. Then run only the command below that matches your selected agent to copy the task skill into its skills directory:

# Claude Code
cp -r ../FAST-Evolve/benchmarks/skills/circle-packing .claude/skills/
# OpenCode
cp -r ../FAST-Evolve/benchmarks/skills/circle-packing .opencode/skills/

3. Start Evolution

evotrias

Enter the following in the agent session:

/evolve evolve circle-packing, 2 initial directions, 2 in parallel, 10 tasks

After you confirm the run parameters, EvoTrias explores two initial directions, runs up to two tasks in parallel, and iterates within a budget of 10 tasks. Follow progress at the dashboard URL printed during the run, and review the summary at runs/<run_id>/run_summary.md in the workspace when the run finishes.

To integrate custom optimization tasks or use other coding agents, see the EvoTrias documentation and benchmark task examples.

🏗️ Architecture Overview

EvoTrias separates the evolution process into governance, proposal, and judgment:

Component Responsibility
Structured Governance The controller manages task budgets, scheduling, and search state, using evaluation evidence to select subsequent exploration directions.
Open Proposal Coding agents explore optimization strategies and generate and modify candidate programs.
Protected Judgment Evaluators check candidate validity and compute fitness, providing feedback for selection and the next iteration.

These components collaborate through explicit task and result contracts, forming a loop of proposal → evaluation feedback → selection and iteration.

✨ Features

  • 🔌 Plugin Integration: Integrate evolution capabilities into coding agent workspaces through skills, an MCP server, and hooks.
  • 🧭 Parallel Exploration: Explore multiple initial directions in parallel and continuously optimize within a task budget.
  • 🔄 Evaluation-Driven Search: Select candidates based on validity and fitness, capture useful experience as Notes, and generate strategy guidance when search stagnates.
  • 🛡️ Evaluation Access Control: In Claude Code, permission hooks restrict access to evaluator source code and other tasks' results.
  • 📊 Run Observability: Track progress and review results through a read-only dashboard, run records, and Markdown summaries.

🤝 Contributing

We welcome issues for bug reports and suggestions, and pull requests for code, documentation, and benchmark tasks.

Before contributing, read the openJiuwen community contribution guidelines and complete CLA signing by following the signing guide. Fork this repository, create a branch from the latest main, make and validate your changes, and submit a pull request.

License

This project is licensed under the Apache License 2.0.

This product serves solely as a workflow orchestration tool and does not embed any AI model capabilities. When users integrate AI models for specific business scenarios, they shall bear full responsibility for compliance obligations under the EU AI Act and other relevant regulatory frameworks.


Developed by Taylor Lab, Huawei Technologies Co., Ltd.

Introduction

FAST-Evolve 是面向算法优化的一站式智能体进化工具链。给出算法问题描述和评测标准,它就能驱动 Coding Agent 自主探索算法方案设计,依据评测反馈持续迭代,让更优解不断涌现。

Customize your domain