An open-source framework provides reusable environments, training recipes, and evaluation tools to help developers build and scale autonomous AI agents across multiple application domains.
Microsoft Research has introduced Orchard, an open-source framework designed to simplify the training of autonomous AI agents for real-world tasks. Unlike many existing frameworks that mainly focus on orchestrating or evaluating agents, Orchard provides infrastructure for scalable agent training, enabling developers to create, test, and refine agents across diverse application domains.
At the core of this framework is Orchard Env, an environment service designed for lightweight management of sandbox lifecycles and the reuse of execution environments. With the decoupling of environment handling and training processes, developers can reuse the same infrastructure for various purposes with support for multiple agent harnesses and modelling pipelines.
The framework contains three unique training recipes designed for coding assistants, GUI navigation, and personal assistants. The recipes use a combination of supervised fine-tuning, reinforcement learning, and data distillation to boost performance and reduce data usage and resource requirements. According to Microsoft Research, models developed using this framework are able to achieve competitive results on various public agent benchmarks.
The architecture of Orchard enables developers to employ different language models, training processes, and evaluation procedures without having to modify the underlying infrastructure. In addition, Orchard makes available some reusable datasets, training pipelines, and evaluation tools that can be adapted for new agent building tasks.
By combining reusable execution environments and scalable training and evaluation pipelines, Orchard serves as a platform for building autonomous agents. It exemplifies how open-source infrastructure can serve as a basis for reproducible research as well as rapid prototyping of agentic AI systems.
