Training agents directly on Kubernetes: a walkthrough of Microsoft's Orchard framework

Microsoft Research released Orchard, an open framework that trains agents inside the same harness they are deployed with, on August 3, 2026. The core piece is Orchard Env, a lightweight Kubernetes environment that creates, manages, and removes thousands of isolated components in parallel so that training and deployment share one harness. Orchard-SWE, trained this way, reaches 69.7% on SWE-bench Verified with roughly 3 billion active parameters (30B-A3B), rising to 73.0% with value-model reranking. ASAP works from the Microsoft Research blog directly to lay out Orchard's structure, its numbers, and the questions it leaves open.

Orchard Env ties training and deployment to the same harness

Orchard is built around what Microsoft Research calls the train-deploy mismatch, a gap between the environment an agent trains in and the harness it actually runs in. Agents today run inside stateful, multi-process harnesses that touch file systems, networks, and several processes at once, and open training tools usually cannot handle those harnesses. Researchers therefore train on a simplified stand-in and deploy into the real setting, so the world the policy learned differs from the world it meets in production.

Orchard Env closes that gap at the environment layer. It is a lightweight Kubernetes environment that creates, manages, and removes thousands of isolated components in parallel, providing sandbox management, file I/O, networking, and REST APIs, and it plugs into real agent harnesses such as Codex, OpenClaw, and ZeroClaw. Because training runs end to end inside the harness the agent will ship with, the learned policy sees the same tool-call sequences and the same failure modes it will face at deployment.

How 3 billion active parameters reach 69.7% on SWE-bench Verified

Orchard-SWE is a 30B-A3B model with about 3 billion active parameters, and it scores 69.7% on SWE-bench Verified. Adding a 4-billion-parameter value model trained on past rollouts to rerank candidates lifts that to 73.0%, which Microsoft Research describes as approaching frontier systems that use models more than 10 times larger.

The recipe behind that number stacks several stages. Distillation comes first, using 107,000 agent interactions collected from two advanced open-weight models, MiniMax-M2.5 and Qwen3.5-397B. Credit-assignment supervised fine-tuning follows, designed to learn from partial attempts rather than only complete successes, and reinforcement learning with Balanced Adaptive Rollout targets the sparse success signals that agent tasks produce. On-policy distillation and a process reward model turn those rare terminal rewards into denser training signal.

Web navigation and personal assistance round out the suite

Orchard-GUI handles web navigation as a 4-billion-parameter vision-language model, scoring 74.1% on WebVoyager, 67.0% on Online-Mind2Web, and 64.0% on DeepShop for an average of 68.4%. Its training data consists of 400 distilled demonstrations combined with 2,200 open-ended training tasks. Because the three benchmarks cover general web navigation, live real-world tasks, and shopping scenarios respectively, the result spans a distribution that single-domain overfitting does not explain neatly.

Orchard-Claw, the personal-assistant model, was trained on just 200 synthetic tasks and completes 59.6% of Claw-Eval tasks when given up to three attempts. Pairing the same model with the stronger ZeroClaw agent system raises that to 73.9%. Two hundred training tasks is an unusually small budget for this area, and Microsoft Research offers it as evidence that a faithful environment and harness let you cut task counts sharply.

The bottleneck moved from algorithms to environments

The most consequential part of this release is what got open-sourced rather than what got benchmarked. Microsoft Research published the environment layer itself alongside models and datasets, and that choice encodes a diagnosis: the binding constraint in agent training is infrastructure, not learning algorithms. Public assets in agent research over the past two years clustered around datasets, benchmarks, and prompting strategies, while the execution infrastructure that spins up thousands of sandboxes, tracks their state, and recovers from failure stayed something each organization rebuilt internally.

Packing three domains into one framework points the same way. Software engineering, web navigation, and personal assistance differ in observation format and in reward structure, so training all three on one Orchard Env implies the environment layer is abstracted without domain coupling, which in turn implies a low cost to attach a fourth domain. The framework's value sits there rather than in any single model score.

What impresses and what deserves caution in the numbers

The efficiency per parameter is the striking part. Reaching 69.7% on SWE-bench Verified with 3 billion active parameters puts a small model into a band that much larger models occupied on the same benchmark. The data budgets are small too: 107,000 interactions for software engineering, roughly 2,600 items for web navigation, and 200 tasks for personal assistance say that spending can shift from bulk data collection toward harness fidelity.

The conditions attached to each figure deserve care. The 73.0% includes value-model reranking, which means generating multiple candidates at inference and reranking them with a 4-billion-parameter model, so it carries extra inference compute. The 3-billion active-parameter framing comes from a 30B-scale mixture-of-experts model, so memory footprint should not be computed from the active count. Orchard-Claw's 59.6% assumes up to three attempts, and its 73.9% reflects pairing with a stronger agent system rather than an improvement in the model itself. Strip the conditions from any of the three numbers and the story changes.

What actually changes for teams building on this

For a team adopting agents, Orchard changes an infrastructure decision more than a model decision. Organizations wanting to train agents on their own data have generally picked one of two paths: layer prompts and tools onto a frontier API model, or fine-tune an open-weight model in a simplified environment. The first carries cost and data-egress problems, and the second inherits the train-deploy mismatch outright.

Orchard offers a third path, which is reusing the Kubernetes cluster an organization already runs as the training environment. A model with 3 billion active parameters brings on-premises serving into realistic range, and it makes configurations where data never leaves the organization workable for regulated industries. Running thousands of sandboxes in parallel demands cluster operations skill rather than raw GPU budget, so the barrier to this path lands on platform engineering rather than on research headcount.

Open questions

All three Orchard domains are built on tasks with verifiable success signals, which is what makes their reward definitions tractable in the first place. Software engineering rewards can key off test passes and web navigation off goal completion. Whether the same procedure holds for work where success is ambiguous, such as document drafting or decision support, falls outside what this release demonstrates.

A second gap remains between benchmark scores and real repositories. SWE-bench Verified is a curated task set, and private build environments, internal dependencies, and undocumented conventions are exactly the conditions a benchmark does not reproduce.

Finally, Microsoft Research names reusing training experience toward cumulative agent learning as future work. What shipped is three models trained separately per domain, and whether experience from one domain transfers to another has no data behind it yet. That transfer is the real test of the framework.

Source: ASAP walkthrough based on the Microsoft Research blog post "Orchard: An open framework for scalable agentic AI" (Baolin Peng, Wenlin Yao, Qianhui Wu, Hao Cheng, Jianfeng Gao, August 3, 2026). Code is published at github.com/microsoft/Orchard and datasets at huggingface.co/datasets/microsoft/Orchard.