Microsoft Research has introduced Orchard, an open-source framework built to support the training and evaluation of agents across software engineering, web navigation and personal-assistant tasks. The project pairs reusable environments with released models, data and evaluation methods, and reports competitive SWE-b...
Microsoft Research has released Orchard, an open-source framework intended to make research on scalable agentic AI more reusable and cost-effective across task domains.
At the center of the project is Orchard Env, a reusable environment service for training and evaluating AI agents. According to Microsoft Research, the framework is designed to support several classes of agentic work from the same underlying infrastructure, including software engineering, web navigation and personal-assistant tasks.
That matters because such agents are evaluated in environments where they must complete multistep actions rather than simply produce a text response. Microsoft Research describes examples including fixing bugs in complex codebases, navigating the web on a user’s behalf, and handling workflows involving calendars and email.
The accompanying Orchard paper describes the environment layer as harness-agnostic. Microsoft Research says Orchard can train agents inside real deployment harnesses including Codex, OpenClaw and ZeroClaw, allowing work across task types to draw on shared environments and evaluation methods.
The project includes three task-focused efforts: Orchard-SWE for software engineering, Orchard-GUI for graphical-user-interface navigation, and Orchard-Claw for personal-assistant agents. The arXiv paper characterizes these as three agentic-modeling recipes built around Orchard Env.
Microsoft Research is releasing models and workflows alongside training data and evaluation approaches. The stated goal is to give researchers materials for building and studying open agentic systems, rather than limiting the framework to a single application area.
Microsoft Research reports that Orchard-SWE achieved 69.7% on SWE-bench Verified, rising to 73.0% with value-model reranking. The organization says the system uses about 3 billion active parameters and that its result approaches frontier systems that use models more than ten times larger.
SWE-bench Verified is a benchmark focused on resolving real software issues from code repositories. The reported result is therefore a specific measurement for Orchard-SWE’s software-engineering configuration, not a general measure of performance across every kind of agent task.
Orchard’s broader contribution is its attempt to make the environments around agents reusable. By tying software, GUI and personal-assistant research to a common environment service while publishing associated artifacts, Microsoft Research is positioning the project as an open foundation for testing how smaller open-weight models perform on practical, multistep tasks.
The framework and its reported results are described in Microsoft Research’s Orchard announcement and publication, as well as the related arXiv paper.
Microsoft Research has released Orchard, an open source framework intended to make research on scalable agentic AI more reusable and cost effective across task domains.
A shared foundation for agent research At the center of the project is Orchard Env , a reusable environment service for training and evaluating AI agents.
According to Microsoft Research, the framework is designed to support several classes of agentic work from the same underlying infrastructure, including software engineering, web navigation and personal assistant tasks.
Continue reading
