Quick answer
Microsoft Orchard is an open-source agentic modeling framework built around a shared environment service called Orchard Env. Microsoft Research uses that substrate across three released recipes:
- Orchard-SWE for software-engineering tasks;
- Orchard-GUI for browser-navigation tasks;
- Orchard-Claw for personal-assistant workflows.
The important design choice is separation: the environment runs as a reusable service instead of being embedded in one trainer or agent harness. Data collection, supervised fine-tuning, reinforcement-learning rollouts, and evaluation can therefore use the same sandbox contract.
Orchard is research infrastructure, not a production-readiness certificate. Reported benchmark results do not establish that an agent is safe, authorized, reliable, or economical in your environment.
The three layers
| Layer | Role | Start here |
|---|---|---|
| Orchard Env | Creates isolated Kubernetes sandboxes and exposes command, file, patch, and lifecycle operations | Orchard Env guide |
| Domain recipes | Demonstrate training and evaluation for code, browser, and assistant tasks | SWE, GUI, or Claw |
| Trainer and harness integration | Runs SFT/RL work and can collect trajectories from real agent harnesses | Harness guide |
The repository also publishes datasets and evaluation artifacts. Treat these as versioned research inputs: inspect provenance, licenses, task overlap, failure labels, and benchmark rules before using them.
Choose Orchard by the bottleneck
Choose Orchard Env when your main problem is repeatedly building isolated execution infrastructure for different agent projects. Evaluate the SWE recipe when the decision concerns repository repair, the GUI recipe for live browser work, and Claw for productivity tools such as email or calendars.
If a team already has an agent harness but training uses a simplified imitation, focus on the real-harness training boundary. If the problem is learning from partial or sparse-reward trajectories, read the credit-assignment guide.
Adoption checklist
Before a pilot, write down:
- the exact task distribution and success measure;
- the environment image, network policy, resource limits, and cleanup owner;
- the model, inference service, trainer, and harness versions;
- which data may enter trajectories and who can inspect them;
- the offline evaluation suite and production approval gate;
- how failed or interrupted rollouts are attributed;
- the cost ceiling and rollback path.
Run an isolated evaluation first. Reproduce a documented baseline on your own cluster, then add one domain-specific change at a time. Do not infer production performance from a paper score alone.
Frequently asked questions
What is Microsoft Orchard?
Orchard is an open-source framework for agentic modeling research. It combines the reusable Kubernetes-native Orchard Env service with domain recipes for software engineering, browser navigation, and personal-assistant work.
Is Orchard an agent application or a training framework?
It is primarily research infrastructure and a collection of training recipes, datasets, and evaluation methods. It does not by itself define your production permissions, safety policy, model endpoint, or deployment approval.
Which Orchard guide should I read first?
Start with Orchard Env if infrastructure is the main decision. Use the SWE, GUI, or Claw guide for a domain recipe, the harness guide for train-deploy alignment, and the credit-assignment guide for the SFT and RL methods.
Official sources
Source check: August 4, 2026. Verify repository state, model and harness versions, datasets, infrastructure requirements, benchmark rules, and licenses before use.