Microsoft Research Introduces Orchard, an Open Framework for Training and Evaluating AI Agents

A reusable foundation for agent research

Microsoft Research has introduced Orchard, an open-source framework intended to make it easier for researchers to train and evaluate AI agents across different types of tasks. According to the Microsoft Research announcement, the project aims to reduce the engineering complexity involved in agentic AI research while allowing teams to reuse a common foundation.

AI agents typically need more than a language model alone: they must operate within task environments, use tools, receive feedback, and be assessed against consistent evaluation criteria. Building those components separately for each experiment can make results harder to compare and can require substantial implementation work. Orchard is positioned as a framework for addressing that repeated infrastructure burden.

An environment layer designed for reuse

The accompanying Orchard paper on arXiv describes the project as an open-source, scalable agentic-modeling framework. Its central component is an Orchard Env service, which the authors characterize as reusable and harness-agnostic. In practical terms, that means the environment layer is meant to work independently of a particular agent harness or task-specific implementation.

The paper says Orchard includes recipes for agents operating in coding, graphical user interface, and personal-assistant settings. These task categories require different interactions with their environments, but the framework is designed to provide a shared basis for training and evaluation rather than requiring researchers to construct a separate system for every domain.

Microsoft Research says this reuse can lower complexity while supporting capable results from smaller models. The announcement does not present Orchard as a replacement for model development; instead, it focuses on making the surrounding experimentation and evaluation infrastructure more accessible and repeatable.

Dataset release supports reproducibility

Microsoft has also published an accompanying Orchard dataset on Hugging Face. The dataset is identified as a trajectory dataset for the Orchard paper. Its description states that the agentic-modeling data was produced with the same Orchard Env sandbox infrastructure described by the project.

That connection between the framework and the released data is significant for researchers seeking to inspect how agent trajectories were produced or to use comparable environment conditions in later work. Shared environments and released trajectories can help separate changes in an agent model or training method from changes introduced by the underlying task setup.

A focus on comparable agent experiments

Orchard arrives as agent research increasingly spans software engineering, computer interaction, and assistant-style workflows. Those settings often differ in tools, state management, and evaluation methods, which can make direct comparison difficult.

Microsoft Research's proposal is to provide common infrastructure that can be applied across those settings. The project's usefulness will depend on how broadly its environment service and task recipes are adopted, as well as on whether external researchers can extend them to additional domains. For now, Orchard offers an open framework, a research paper, and a related trajectory dataset aimed at making scalable agent training and evaluation easier to reproduce and reuse.