Orchard

Orchard is an open-source framework for training LLM-based agents across diverse task domains. It provides reusable agentic data, training recipes, and evaluations — covering coding, GUI navigation, and personal assistant tasks — through a harness-agnostic design that separates environment management from training logic.

Orchard Env provides lightweight sandbox lifecycle management. Three specialized recipes demonstrate the framework’s reach:

• Orchard-SWE (coding agents): 67.5% on SWE-bench Verified with Qwen3-30B via SFT+RL
• Orchard-GUI (vision-language agents): 74.1% on WebVoyager with a 4B model
• Orchard-Claw (personal assistants): 73.9% pass@3 on Claw-Eval