Orchard: An open framework for scalable agentic AI

Orchard is an open-source framework designed to facilitate scalable and cost-effective research in agentic AI. At its core lies Orchard Env, a reusable environment service that enables training and evaluation of agents across various task domains. This framework supports a wide range of agents, including those for software engineering, web navigation, and personal assistance, and can be utilized directly within real-world deployment environments such as Codex, OpenClaw, and ZeroClaw.

This allows researchers to reuse environments, data pipelines, and evaluation workflows across different tasks.

Orchard-SWE, Orchard-GUI, and Orchard-Claw are examples of domain-specific training recipes that demonstrate the framework's capabilities. For instance, Orchard-SWE achieves impressive results on the SWE-bench Verified benchmark, reaching 69.7% accuracy, which improves to 73.0% when incorporating value-model reranking. Remarkably, this is accomplished using only about 3 billion active parameters, a level that approaches that of frontier systems typically requiring models with over 10 times more parameters.

The project also emphasizes the importance of releasing training data and evaluation methods to foster a collaborative research environment focused on developing and studying open agentic systems. As artificial intelligence advances beyond static question-answering towards autonomous agents capable of planning, reasoning, and acting within complex, multistep environments, the demand for accessible infrastructure to support these systems grows.

Orchard addresses this need by providing an open-source solution that bypasses proprietary infrastructure, including custom sandboxes and closed training pipelines.