The Brief: Microsoft has introduced Orchard, an open-source framework for research and development of agentic AI systems. The framework is built around Orchard Env, a lightweight Kubernetes environment service providing reusable isolated components for agent training, data collection, reinforcement learning rollouts, and evaluation.
The company is releasing three training recipes: Orchard-SWE for software engineering, Orchard-GUI for browser navigation, and Orchard-Claw for personal-assistant tasks. The project also includes training data and evaluation methods.
Microsoft reports that Orchard-SWE reaches 69.7% on SWE-bench Verified, rising to 73% with value-model reranking, while using about 3 billion active parameters. Orchard-GUI records a 68.4% average across three web-navigation benchmarks. Orchard-Claw reaches 59.6% on Claw-Eval with up to three attempts.
Analyst Perspective: Microsoft’s Orchard initiative adds an infrastructure layer to its growing body of agentic AI research. The distinction is important because the project is not limited to releasing another trained model. Orchard Env provides a common runtime that can support data generation, reinforcement learning, and evaluation across several agent categories. This gives Microsoft a way to connect research workflows that would otherwise require separate environments.
The support for established agent harnesses also gives Orchard relevance for developers working with increasingly sophisticated agent systems. Codex, OpenClaw, and ZeroClaw can contain multiple processes, tool calls, and persistent interactions that are difficult to reproduce in simplified training loops. Orchard’s ability to operate within these environments brings training activity closer to the conditions under which agents are actually used.
Microsoft is also contributing datasets, recipes, and evaluation methods. That combination gives researchers more material for reproducing experiments and developing additional open agentic systems. The emphasis on reusable infrastructure makes Orchard particularly relevant to organizations building research programs across multiple agent workloads.
Orchard Env is designed as a standalone service that separates the execution environment from individual training frameworks. It’s built on Kubernetes and can create and manage isolated components for agent execution, including command execution, file access, network controls, sandbox management, and agent integration through a REST API.
This design allows the same environment service to support different training stages and task categories. Researchers can use it for data distillation, reinforcement learning rollouts, and model evaluation without creating separate infrastructure for each activity. New benchmarks or agent configurations can also be introduced without rebuilding the underlying environment.
Another capability is direct training through deployment harnesses. Orchard uses a lightweight proxy to record model calls generated within a harness while each rollout operates inside its own container. This allows training activity to capture the multi-turn interactions and tool usage associated with systems such as Codex, OpenClaw, and ZeroClaw.
Orchard-SWE applies the framework to software engineering through Mini-SWE-Agent and the SWE-bench Verified benchmark. The training process uses 107,000 agent interactions distilled from MiniMax-M2.5 and Qwen3.5-397B across a wide collection of GitHub issues. Microsoft’s method retains useful portions of incomplete attempts through credit-assignment supervised fine-tuning, increasing the amount of usable training material.
The reinforcement learning process adds several forms of feedback. Balanced Adaptive Rollout is used to work with sparse success signals, while on-policy distillation and a process reward model provide additional guidance on individual decisions. The latter can reward behaviors such as reproducing bugs, testing fixes, and checking that existing functionality remains intact.
Orchard-SWE improved from a 61.4% baseline to 69.1% with Balanced Adaptive Rollout and 69.7% after adding dense-reward techniques. Value-model reranking raises performance to 73%, using a 4-billion-parameter value model trained from previous experiments.
Orchard-GUI applies Orchard to browser-based work through a 4-billion-parameter vision-language model. Its training uses 400 distilled demonstrations and 2,200 open-ended tasks. The model records 74.1% on WebVoyager, 67.0% on Online-Mind2Web, and 64.0% on DeepShop, producing an average score of 68.4% across the three benchmarks.
Orchard-Claw applies the same infrastructure to personal-assistant activities such as email handling, calendar management, information retrieval, and coordination across tools. It uses 200 synthetic tasks and achieves a 59.6% success rate on Claw-Eval when given up to three attempts. With ZeroClaw, performance reaches 73.9%.
Training across ReACT, ZeroClaw, OpenClaw, and Codex also provides measurable differences in performance. Under Codex, the trained agent improves from 18.6% for the untrained model to 51.5% after Orchard training. These examples show how the framework can support agents across different interaction models.
Microsoft’s Orchard initiative adds to the company’s ongoing work in agentic AI, developer tools, and open research.
By providing reusable environments for training and testing AI agents, Orchard gives researchers and developers another way to build and experiment with these systems. It also complements Microsoft’s work across AI models, coding tools, cloud infrastructure, and enterprise software.
For teams developing different types of agents, shared environments and training resources could reduce the need to build separate infrastructure for every project. This could be especially useful for organizations working on several agent applications that require reliable tools for training and testing.
Open infrastructure does not remove the technical demands of agent development. Teams still need reliable benchmarks, high-quality trajectories, compute resources, and methods for controlling agent behavior.
Compatibility with different harnesses can also require additional engineering. Clear documentation, reproducible configurations, active community contributions, and continued dataset releases can help address these issues.
Microsoft’s work with Orchard could make it easier to preserve useful training experiences and use them to improve future agents. The company’s use of previous training trajectories to build a value model already offers an example of how past agent experiences can remain useful after a training run ends.
If Orchard gains wider adoption among researchers and developers, its shared environment, datasets, and training recipes could provide a solid foundation for creating and testing more capable open agent systems.
Make confident decisions with research and advisory built for technology providers. Contact us at the Collab Collective to see how tailored insights can guide your next strategy.