Orchard - an open-source agentic AI modeling framework from Microsoft Research. - AiBoss

What is Orchard?

Orchard is an open-source agentic AI modeling framework developed by Microsoft Research. Its core is the Orchard Env environment service based on Kubernetes, supporting cross-domain reuse of sandboxes for data distillation, reinforcement learning rollout, and evaluation. The framework covers three major scenarios: software engineering, browser navigation, and personal assistants, and has open-sourced its training data and evaluation methods.

Orchard's main functions

Cross-domain unified environment services: Orchard Env provides sandbox lifecycle management, command execution, file I/O, network policies, and REST APIs, offering a single infrastructure that supports code, web, desktop, mobile, and productivity workflows.
Training formulas for three major domains: Orchard-SWE specializes in code repair, Orchard-GUI handles visual web page navigation, and Orchard-Claw is geared towards everyday productivity tasks such as email and calendar, all providing complete SFT + RL training pipelines.
Real Harness training: It supports end-to-end training and evaluation directly within real-world deployment harnesses such as Codex, OpenClaw, and ZeroClaw, eliminating mismatches between training and deployment environments.
Open source datasets and evaluation protocols: Release 107K SWE trajectories and 3,070 GUI multimodal rollout data, along with corresponding benchmarks and reproducible recipes.

Orchard's technical principles

Kubernetes native environment layer: Orchard Env runs as a standalone service, enabling the parallel creation and destruction of thousands of isolated sandboxes through container orchestration, with an average command execution latency of only 0.28 seconds. It supports automatic scaling and Redis distributed locks.
Credit-Assignment SFT： In the field of software engineering, instead of discarding partially failed trajectories, the system extracts supervisory signals from the effective fragments to maximize the amount of available training data.
Balanced Adaptive Rollout + Dense Bonus: To address the sparse feedback problem in RL, a balanced adaptive rollout is employed to capture sparse success signals, and on-policy distillation and a rule-based process reward model are introduced to provide dense guidance for intermediate steps.
Value Model Reordering: A 4B parameter value model was trained using the rollout trajectory of 20 historical experiments. During the inference phase, multiple candidate solutions were scored and the optimal solution was selected, which improved the Orchard-SWE from 69.7% to 73.0%.
Harness-Agnostic proxy history: The lightweight proxy records the harness's own model calls and reconstructs them into training samples, enabling any RL codebase to interface with any harness without modifying the environment image.

How to use Orchard

Install SDK: implement pip install -e "orchard_env[dev]" Install the Python SDK and configure it. SANDBOX_BASE_URL and SANDBOX_API_KEY Environment variables.
Quickly invoke the sandbox: use SandboxClient The context manager creates sandboxes and executes commands, supporting synchronous and asynchronous modes, file reading and writing, git patch application, and PTY sessions.
Deploy orchestrator: The multi-replica orchestrator can be deployed in about 20 minutes using the four scripts (provision, build, deploy, smoke-test) provided by Azure AKS, and it also supports non-Azure clusters.
Run the training recipe: Refer to the GitHub repository trainer/slime/ This document outlines the complete SFT + RL training process for Orchard-SWE, Orchard-GUI, or Orchard-Claw on Orchard Env.

Orchard's core advantages

Smaller models approximate the performance of larger models: Orchard-SWE achieved a score of 73.0% on SWE-bench Verified with approximately 3B active parameters, approaching that of state-of-the-art systems with more than 10 times the number of parameters; Orchard-GUI achieved an average score of 68.4% on the three major web benchmarks with 4B parameters, comparable to OpenAI CUA and Gemini CUA.
Costs are significantly reduced: Compared to managed sandbox services such as E2B and Daytona, Orchard Env's on-demand cost is approximately 0.47 times that of Spot instances, while Spot instances cost only 0.10 times, representing a price difference of about 10 times.
Strong generalization ability across Harness: The same environment layer supports multiple harnesses such as ReACT, ZeroClaw, OpenClaw, and Codex. Orchard-Claw's pass@3 rate under ZeroClaw increased from 59.6% to 73.9%.
Training data is efficient: Orchard-GUI achieved the level of the most powerful open-source web agent with only 400 distillation demos and 2,200 open tasks, proving that a high-performance model can be trained with a small amount of data and the right environment.

Orchard's project address

Project 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
HuggingFace model library：https://huggingface.co/datasets/microsoft/Orchard
arXiv technical paper：https://arxiv.org/pdf/2605.15040

Comparison of Orchard's similar products

Dimension	Orchard	OpenHands
position	Microsoft's open-source cross-domain Agentic modeling framework covers three major scenarios: code, web pages, and productivity.	Open-source community-led software engineering agents, focusing on codebase fixes and development tasks.
Environment layer	We have developed a self-developed Kubernetes-native general-purpose environment service (Orchard Env), enabling cross-domain reuse of a single infrastructure.	Relying on Docker containerized environments and customized for code tasks, it is difficult to directly migrate to web or desktop scenarios.
Training ability	It includes a complete SFT + RL training pipeline, supporting direct training within real-world harnesses such as Codex and OpenClaw.	It primarily provides the inference and execution framework; the training pipeline needs to be built independently and is typically performed in a simplified environment.
Data open source	Open source: 107K SWE trajectories, 3,070 GUI multimodal rollouts, and a complete evaluation protocol.	Open source code and some evaluation results are available, but the large-scale training trajectory dataset is not fully disclosed.
Model efficiency	With 3B–35B active parameters, a score of 73.0% can be achieved in SWE-bench, which is close to that of a frontier model with 10 times more parameters.	Similar SWE-bench scores are typically achieved only with 70B-level models or commercial APIs such as GPT-4.
Cost structure	Self-hosted sandboxes cost approximately 10%–50% of commercial services, while supporting Spot instances significantly reduces training overhead.	Primarily reliant on cloud servers or commercial APIs, costs increase linearly with model size and call volume.
Harness Support	Natively supports seamless switching between multiple harnesses such as ReACT, ZeroClaw, OpenClaw, and Codex.	The design primarily revolves around its own harness; switching to an external harness requires additional adaptation.

Application scenarios of Orchard

Automated Software Engineering: It enables users to independently diagnose bugs, write tests, generate patches, and pass verification in real codebases, making it suitable for open-source project maintenance and internal enterprise code review.
Intelligent webpage navigation: Visual understanding-based automatic browser operation to complete open domain tasks such as ticket booking, shopping, and information retrieval can serve as an alternative to RPA.
Personal productivity assistant: Execute complex workflows across tools such as email, calendar, and documents, such as automatically scheduling meetings, organizing inboxes, and generating reports.
Agent research infrastructure: It provides academic and industrial research teams with a reproducible, low-cost agent training and evaluation platform, accelerating the development of the open-source agentic AI ecosystem.
Cross-domain Agent capability migration: Explore skill transfer and cumulative learning across code, web pages, and desktop operations using a unified environment layer.