Microsoft Research published Orchard on August 3, an open-source framework aimed at a bottleneck the company says has been slowing agentic AI research: most state-of-the-art agent systems are trained on proprietary infrastructure, custom sandboxes and closed datasets that other teams can't reuse. Orchard's core piece, Orchard Env, is a lightweight, Kubernetes-based environment service that gives researchers reusable, isolated components for running and training agents, rather than requiring them to rebuild task environments from scratch for every new domain. Both the framework and its training data and evaluation methods are released publicly, on GitHub and Hugging Face respectively.

Three ready-made training recipes

Alongside the base framework, Microsoft shipped three domain-specific recipes that show what Orchard can produce. Orchard-SWE trains agents to fix bugs and solve coding tasks in real repositories, evaluated on the SWE-bench Verified benchmark. Orchard-GUI trains vision-language-model-based browser agents to interpret on-screen layouts and complete web tasks. Orchard-Claw targets personal-assistant workflows such as email triage, calendar coordination and multi-tool task completion.

Performance claims

Microsoft reports Orchard-SWE reaches 69.7% on SWE-bench Verified (73% with value-model reranking) using an agent with roughly 3 billion active parameters — a result the company says approaches systems more than ten times larger. Orchard-GUI is reported to average 68.4% success across three web-navigation benchmarks, in the range of proprietary browser agents from OpenAI and Google. Orchard-Claw completes 59.6% of assigned tasks within three attempts on its own, rising to 73.9% when paired with a stronger downstream agent system.

Why it matters

Orchard is a bet that the next constraint on agent capability isn't model size but training infrastructure: giving smaller, cheaper models realistic environments to practice in before deployment, rather than routing everything through ever-larger general-purpose models. By open-sourcing the environment layer, benchmark recipes and training data together, Microsoft is inviting outside teams to build and compare their own agents on the same substrate — a departure from labs that keep their agent-training pipelines closed. If independent teams can reproduce Orchard's reported results, it could lower the cost of entry for building specialized coding, browser and assistant agents without needing frontier-scale compute budgets.