Microsoft Research has introduced Orchard, an open-source framework for training and evaluating AI agents across software engineering, web navigation and personal-assistant tasks. The release includes reusable environments, model-training recipes, data and evaluation methods.
Microsoft Research has released Orchard, an open-source framework intended to make agentic AI research more scalable and cost-effective.
The project is built around Orchard Env, a reusable environment service for training and evaluating AI agents across different task domains. According to the Microsoft Research blog, the framework is designed to let researchers reuse environments, training data and evaluation methods rather than creating separate technical foundations for each agent application.
Microsoft Research says Orchard supports software-engineering, web-navigation and personal-assistant agents. It can also train agents within deployment harnesses including Codex, OpenClaw and ZeroClaw.
Agentic systems are intended to perform multistep tasks rather than simply return an answer to a prompt. Those tasks can include modifying code in a repository, navigating websites, or handling work involving calendars and email.
Microsoft Research argues that research in this area has been constrained by environments that are often built for a single benchmark or use case. Orchard’s environment layer is described as harness-agnostic, meaning researchers can use it with different agent configurations and deployment setups.
Microsoft’s public GitHub repository describes Orchard as an open foundation for agentic modeling. It includes a Kubernetes-native Orchard Env sandbox service as well as reusable supervised fine-tuning and reinforcement-learning recipes. A related Orchard paper on arXiv similarly describes a framework with recipes for software engineering, graphical-interface and personal-assistant agents.
The release includes example systems named Orchard-SWE, Orchard-GUI and Orchard-Claw. Microsoft Research presents them as evidence that relatively small open-weight models can deliver strong results on complex agent tasks.
For software engineering, Microsoft Research reports that Orchard-SWE scored 69.7% on SWE-bench Verified, a benchmark based on resolving software issues in real code repositories. The reported score increased to 73.0% when using value-model reranking. Microsoft says the system used about 3 billion active parameters, while approaching the reported performance of frontier systems using models more than 10 times larger.
These are benchmark results reported by the project’s authors. SWE-bench Verified measures performance on a defined set of repository issues and does not by itself establish reliability across broader production codebases, toolchains or operational conditions.
Microsoft Research says Orchard also releases training data and evaluation methods for the research community. The stated aim is to help outside developers and researchers build, test and compare open agentic systems across several task types.
The practical importance of Orchard will depend on whether researchers can reproduce its published results, extend its environments to additional domains and use its components to make comparisons between agent-training approaches more consistent.
Microsoft Research introduces Orchard Microsoft Research has released Orchard , an open source framework intended to make agentic AI research more scalable and cost effective.
The project is built around Orchard Env , a reusable environment service for training and evaluating AI agents across different task domains.
Microsoft Research says Orchard supports software engineering, web navigation and personal assistant agents.
Continue reading
