Microsoft Research has released Orchard, an open source framework designed to make training autonomous AI agents easier and more cost-effective. The framework was released on Monday, August 4, 2026, building on the company's initial research published in March. The release arrives as agentic AI — systems that can autonomously plan, browse the web, write code, and use tools — has become one of the most active areas of industry competition, with major labs and startups alike racing to build agents that can handle multi-step real-world tasks.

Orchard is capable of training and evaluating AI agents across multiple domains, including coding, web browsing, and tool usage. It aims to reduce complexity by eliminating the need for researchers to independently build sandbox infrastructure, data pipelines, and evaluation systems for each individual model or use case — a barrier that has historically put cutting-edge agentic research out of reach for many academic and independent teams lacking the engineering resources of large technology companies.

In a blog post, Microsoft acknowledged the challenges facing the research community: "While there is excitement around agentic AI's capabilities, the research community faces a persistent bottleneck. Building state-of-the-art agentic systems often requires proprietary infrastructure … that most researchers and practitioners cannot access or reproduce."

At the center of the framework is Orchard Env, which Microsoft describes as a "lightweight, Kubernetes environment that provides reusable isolated components." These components can be used to run and build agents at scale — from collecting training data to reinforcement learning rollouts and evaluation — without requiring teams to rebuild the underlying architecture each time. By building on Kubernetes, an industry-standard container orchestration platform, Orchard is designed to be deployable on a range of infrastructure from local clusters to major cloud providers.

To demonstrate the framework's approach, Microsoft released three domain-specific training recipes — Orchard-SWE, Orchard-GUI, and Orchard-Claw — along with the associated training data and evaluation methods used to develop them. Orchard-SWE trains software-engineering agents, Orchard-GUI focuses on browser navigation, and Orchard-Claw addresses everyday productivity tasks. Microsoft reported that models trained through each recipe produced results competitive on established benchmarks.

The company highlighted the broader implications of the framework's open design. "By making the underlying infrastructure open, lightweight, and reusable, Orchard lowers the cost of agentic AI research," Microsoft said. "Teams no longer need to build custom isolated environments from scratch or depend on proprietary cloud services." The open source approach aligns with a broader industry pattern in which major AI labs — including Meta with its Llama models and Google with several research frameworks — have released tools and model weights publicly to accelerate community-driven progress.

Microsoft also outlined its perspective on future development directions. "Looking ahead, we see reusing training experience as a promising direction toward cumulative agent learning. Instead of discarding trajectories once a training run finishes, we treat them as persistent assets — for example, distilling them into reusable value models." This concept of cumulative learning — where agents build on prior experience rather than starting each training run from scratch — mirrors longstanding goals in reinforcement learning research, where sample efficiency and knowledge transfer remain open challenges.

The Orchard framework, along with its associated research publication, is available as an open source release from Microsoft Research.