Microsoft Research Blog
1. Orchard: An open framework for scalable agentic AI
微软的Orchard框架解决了可扩展智能体AI基础设施的关键需求，提供了可复用组件和出色的基准表现。其开源特性和全面的配方使其对构建实际AI智能体的研究人员和从业者具有即时参考价值。 (score: 0.92)
Microsoft Research introduces Orchard, an open-source framework for scalable agentic AI research. At its core is Orchard Env, a Kubernetes-based environment service that provides reusable isolated components for training and evaluating agents across diverse task domains, including software engineering, web navigation, and personal assistants. The framework supports training agents directly inside real deployment harnesses such as Codex, OpenClaw, and ZeroClaw. Three domain-specific recipes are released: Orchard-SWE achieves 69.7% on SWE-bench Verified with about 3 billion active parameters, Orchard-GUI reaches 68.4% average on web-navigation benchmarks, and Orchard-Claw completes 59.6% of productivity tasks. The project also releases training data and evaluation methods to foster open research.
- Orchard is an open framework with Orchard Env, a Kubernetes-based reusable environment service.
- It supports training agents directly inside real deployment harnesses like Codex, OpenClaw, and ZeroClaw.
- Orchard-SWE reaches 69.7% on SWE-bench Verified using ~3B active parameters, approaching larger frontier models.
- Orchard-GUI achieves 68.4% average success on web navigation benchmarks, strong for a 4B-parameter model.
- Orchard-Claw completes 59.6% of personal-assistant tasks, improving to 73.9% with ZeroClaw.
- The project releases training data, evaluation methods, and workflows to support open agent research.