Agent-Field / CodeAF
An open-source Go-binary coding harness that turns your terminal into a parallel agent software factory optimized for open models like DeepSeek and Qwen
星标趋势
AI 分析
项目摘要
CodeAF is an open-source coding harness built specifically for open models (DeepSeek, Qwen, GLM, Kimi, MiniMax, Ollama) that transforms the terminal into a 'software factory' where developers can delegate, monitor, and manage multiple parallel AI coding agents from a single window.
为什么值得关注
It claims #1 on DeepSWE benchmarks among ten coding harnesses using the same model, outperforming Claude Code, Codex, OpenCode, Kilo, and DeepSeek's own harness at the lowest cost per solved issue. Its factory-style approach to orchestrating multiple parallel sub-agents for coding tasks is a distinct take on the coding agent paradigm.
优势
- Top benchmark results (DeepSWE #1) with open models at lower cost than competitors
- Single Go binary with zero dependencies, runs on darwin/linux/windows
- Very active development with 86 releases in 6 months and Apache 2.0 license
- Multi-model support across DeepSeek, Qwen, GLM, Kimi, MiniMax, and Ollama
- Factory UI concept enabling parallel multi-agent management from one window
局限性
- Early preview status with acknowledged rough edges
- 260 open issues suggesting significant stability/feature gaps
- No Docker support and no examples provided
- Relatively small contributor base (9 contributors) for an ambitious project
使用场景
- AI-assisted software development using open-source LLMs instead of proprietary APIs
- Parallel multi-task coding workflows where multiple agents work simultaneously on different features
- Cost-conscious development teams seeking frontier-grade coding assistance without API costs
- Headless and remote coding agent orchestration from terminal or mobile