Agent training has long faced a contradiction: closed-source systems deliver superior results but keep their code proprietary, while open-source frameworks offer flexibility but mostly remain at the orchestration and evaluation stages, lacking scalable training capabilities. Microsoft Research's paper submitted on May 14, Orchard, aims to bridge this gap.
Core Concept: A Lightweight Environment Layer
At the heart of Orchard is Orchard Env—a lightweight environment service that provides reusable sandbox lifecycle management primitives across different task domains. It is not tied to any specific framework and is agnostic enough to support various types of Agent training.
Built on top of this environment layer, Orchard introduces three Agent training recipes, covering Code Agents, GUI Agents, and personal assistants.
Orchard-SWE: 67.5% on SWE-bench Verified
This is the most technically intensive component. The team distilled 107,000 trajectories from MiniMax-M2.5 and Qwen3.5-397B, introducing a technique called credit-assignment SFT—which learns useful intermediate segments from previously unsolved trajectories. This is combined with Balanced Adaptive Rollout for reinforcement learning.
Results: Based on Qwen3-30B-A3B-Thinking, the model achieved 64.3% on SWE-bench Verified after SFT, and reached 67.5% after RL—setting a new SOTA for open-source models of comparable size.
Orchard-GUI: 400 Trajectories Are Enough
A 4B-parameter vision-language model, trained with just 400 distilled trajectories and 2,200 open-ended tasks, achieved 74.1% on WebVoyager and 67.0% on Online-Mind2Web. This marks the strongest performance among open-source models, while remaining competitive with closed-source systems.
Orchard-Claw: Personal Assistant
Trained on only 200 synthetic tasks, it achieved a 59.6% pass@3 score on Claw-Eval. When paired with the ZeroClaw harness, performance jumped to 73.9%.
Significance
Orchard demonstrates that a lightweight, open-source, harness-agnostic environment layer can enable the reuse of Agent data, training recipes, and evaluation metrics across different domains. For teams looking to train Agents without access to closed-source infrastructure, this currently stands as one of the most comprehensive solutions in the open-source community.