Orchard – Microsoft Research’s Open-Source Agentic AI Modeling Framework Operational

Orchard is an open-source Agentic AI modeling framework released by Microsoft Research. Its core component is the Kubernetes-based Orchard Env environment service, which supports cross-domain sandbox reuse for data distillation, reinforcement learning rollouts, and evaluation. The framework covers three major scenarios—software engineering, browser navigation, and personal assistants—and has released its training data and evaluation methods.

Cross-domain unified environment service: Orchard Env provides sandbox lifecycle management, command execution, file I/O, network policies, and a REST API. A single infrastructure can support code, web, desktop, mobile, and productivity workflows.
Training recipes for three major domains: Orchard-SWE focuses on code repair, Orchard-GUI handles visual web navigation, and Orchard-Claw targets everyday productivity tasks such as email and calendar management. All three provide complete SFT + RL training pipelines.
Training in real harnesses: Supports end-to-end training and evaluation directly in real deployed harnesses such as Codex, OpenClaw, and ZeroClaw, eliminating mismatches between training and deployment environments.
Open-source datasets and evaluation protocols: Releases 107K SWE trajectories and 3,070 GUI multimodal rollout data points, along with corresponding evaluation benchmarks and reproducible recipes.

Follow us on WeChat and reply “开源” to join the AI open-source project discussion group.
Kubernetes-native environment layer: Orchard Env runs as an independent service, using container orchestration to create and destroy thousands of isolated sandboxes in parallel. Average command execution latency is only 0.28 seconds, with support for automatic scaling and Redis distributed locks.
Credit-Assignment SFT: In software engineering, the system does not discard partially failed trajectories. Instead, it extracts supervision signals from valid segments to maximize the amount of usable training data.
Balanced Adaptive Rollout + dense rewards: To address sparse RL feedback, Orchard uses balanced adaptive rollouts to capture rare success signals, and introduces on-policy distillation and a rule-based process reward model to provide dense guidance for intermediate steps.
Value Model reranking: Rollout trajectories from 20 historical experiments are used to train a 4B-parameter value model. During inference, it scores multiple candidate solutions and selects the best one, improving Orchard-SWE from 69.7% to 73.0%.
Harness-agnostic agent recording: A lightweight agent-recording harness captures the harness’s own model calls and reconstructs them as training samples, allowing any RL codebase to connect to any harness without modifying the environment image.

Install the SDK: Run pip install -e "orchard_env[dev]" to install the Python SDK, then configure the SANDBOX_BASE_URL and SANDBOX_API_KEY environment variables.
Quickly call a sandbox: Use the SandboxClient context manager to create a sandbox and execute commands. Synchronous and asynchronous modes, file reads and writes, git patch application, and PTY sessions are supported.
Deploy the orchestrator: Using the four scripts provided for Azure AKS—provision, build, deploy, and smoke-test—you can complete a multi-replica orchestrator deployment in approximately 20 minutes. Non-Azure clusters are also supported.
Run training recipes: Refer to the trainer/slime/ directory in the GitHub repository to run the complete SFT + RL training workflows for Orchard-SWE, Orchard-GUI, or Orchard-Claw on Orchard Env.

Smaller models approach the performance of larger models: Orchard-SWE achieves 73.0% on SWE-bench Verified with approximately 3B active parameters, approaching frontier systems with more than 10 times as many parameters. Orchard-GUI achieves an average of 68.4% across three major web benchmarks with 4B parameters, comparable to OpenAI CUA and Gemini CUA.
Significantly lower costs: Compared with managed sandbox services such as E2B and Daytona, Orchard Env’s on-demand cost is approximately 0.47 times as high, while its Spot instance cost is only 0.10 times as high—a roughly 10-fold price difference.
Strong cross-harness generalization: The same environment layer supports multiple harnesses, including ReACT, ZeroClaw, OpenClaw, and Codex. With ZeroClaw, Orchard-Claw improves pass@3 from 59.6% to 73.9%.
Efficient training data use: Orchard-GUI reaches the strongest open-source web-agent performance using only 400 distilled demonstrations and 2,200 open-ended tasks, demonstrating that a small dataset combined with the right environment can train highly capable models.

Official website: https://www.microsoft.com/en-us/research/blog/orchard-an-open-framework-for-scalable-agentic-ai/
GitHub repository: https://github.com/microsoft/Orchard
Hugging Face model repository: https://huggingface.co/datasets/microsoft/Orchard
arXiv technical paper: https://arxiv.org/pdf/2605.15040

Automated software engineering: Autonomously diagnose bugs, write tests, generate patches, and validate them in real codebases, supporting open-source project maintenance and internal enterprise code review.
Intelligent web navigation: Use visual understanding to operate browsers automatically for open-domain tasks such as booking tickets, shopping, and information retrieval, serving as an alternative to RPA solutions.
Personal productivity assistant: Execute complex workflows across email, calendars, documents, and other tools, such as automatically scheduling meetings, organizing inboxes, and generating reports.
Agent research infrastructure: Provide academic and industrial research teams with a reproducible, low-cost platform for Agent training and evaluation, accelerating the development of the open-source Agentic AI ecosystem.
Cross-domain Agent capability transfer: Use a unified environment layer to explore skill transfer and cumulative learning across code, web, and desktop operations.