Microsoft Research has released Orchard, an open-source framework designed to train and evaluate AI agents across software engineering, web navigation, and personal-assistant tasks using a reusable Kubernetes environment service and domain-specific training recipes.
Artificial intelligence is moving beyond static question-answering toward autonomous agents capable of planning, reasoning, and acting across complex, multi-step environments. Yet, the research community has faced a persistent bottleneck: building state-of-the-art agentic systems often requires proprietary infrastructure, custom sandboxes, closed training pipelines, and proprietary datasets that most practitioners cannot access or reproduce. To close this gap, Microsoft Research introduced the Orchard framework as an open-source solution for scalable agentic modeling.
Orchard Env and the Kubernetes Foundation for Agent Training
At the center of the architecture is Orchard Env, a lightweight, Kubernetes-native service that creates and manages isolated components at scale. Rather than embedding runtime infrastructure directly inside a specific training framework, Orchard treats the runtime environment as a standalone, reusable service.
This Kubernetes foundation enables the system to create, manage, and remove thousands of isolated components in parallel. It supports data collection, reinforcement learning rollouts, and evaluation across diverse task types. Developers can reuse the same underlying infrastructure for coding assistants, GUI navigation, and personal assistants without modifying the core system.
By decoupling environment handling from training processes, the architecture allows research teams to introduce new benchmarks, agent systems, or training algorithms without rebuilding foundational infrastructure from scratch.
Training Directly Inside Real Deployment Harnesses
Today’s most capable agents rarely run as bare models. Instead, they operate through sophisticated harnesses—such as Codex, OpenClaw, and ZeroClaw—that manage multi-turn reasoning, tool use, and connections to external systems. Open training tools usually cannot handle these stateful, multi-process harnesses, forcing researchers to train on simplified stand-ins and deploy in real settings, creating a mismatch.
Orchard addresses this mismatch by using a lightweight proxy that records a harness’s own model calls as training data while each rollout runs in a separate container. This setup allows an agent to be trained end-to-end directly inside the harness it will be deployed with.
Domain-Specific Recipes and SWE-Bench Performance
The release provides three distinct domain-specific training recipes: Orchard-SWE, Orchard-GUI, and Orchard-Claw, alongside the underlying training data and evaluation methods.
Orchard-SWE applies the framework to software engineering using Mini-SWE-Agent, evaluated on SWE-bench Verified. The researchers distilled 107,000 agent interactions from open-weight models, utilizing sections of unsuccessful attempts where agents still completed productive work rather than discarding those interactions.
Using approximately three billion active parameters, Orchard-SWE reached a 69.7% score on SWE-bench Verified after applying Balanced Adaptive Rollout and dense-reward techniques, improving from a 61.4% baseline. Performance climbed to 73.0% when candidate solutions were reranked using a four-billion-parameter value model.
Open-Source Availability and Reproducible Research
Alongside the training workflows and models, Microsoft Research has released training data and evaluation methods for the broader research community. By combining scalable execution environments with standardized training and evaluation pipelines, the initiative provides a platform designed to foster reproducible research and rapid prototyping across agentic AI systems.
