Orchard: An open framework for scalable agentic AI
Microsoft Research Baolin Peng, Wenlin Yao, Qianhui Wu, Hao Cheng, Jianfeng Gao
Microsoft Research open-sourced Orchard, a shared training environment for AI agents that do coding, web browsing, and personal-assistant work. Small models trained on it now punch near frontier-level, which is a big deal for anyone without a giant GPU budget.
Microsoft just released something that quietly matters more than another chatbot benchmark: a reusable, open-source environment for training agentic AI, called Orchard. The pitch is simple. Building agents that can code, browse the web, or manage your inbox usually means building custom sandboxes and closed pipelines from scratch every time. Orchard tries to kill that duplication by giving researchers one Kubernetes-based service, Orchard Env, that can spin up thousands of isolated task environments and reuse them across completely different domains.
The proof is in three spinoff projects. Orchard-SWE tackles software engineering and hits 69.7% on SWE-bench Verified, climbing to 73% with a value-model reranking trick, using a model with only about 3 billion active parameters. That's within striking distance of systems ten times its size. To get there, Microsoft distilled 107,000 agent interactions from MiniMax-M2.5 and Qwen3.5-397B, then layered on reinforcement learning with dense reward signals, since raw pass/fail feedback on code fixes is brutally sparse.
Orchard-GUI, a 4-billion-parameter vision-language model, learned to navigate real websites from just 400 demonstrations plus 2,200 open-ended tasks. It averages 68.4% across WebVoyager, Online-Mind2Web, and DeepShop, putting it among the strongest open web agents around, and reportedly competitive with proprietary systems from OpenAI and Google. Orchard-Claw, meanwhile, handles personal-assistant chores like email and calendar juggling after training on a mere 200 synthetic tasks. On the Claw-Eval benchmark it completes 59.6% of tasks solo, jumping to 73.9% when paired with the ZeroClaw harness. Trained inside the Codex harness specifically, its success rate goes from 18.6% untrained to 51.5% after Orchard training.
The common thread across all three is that Orchard trains agents directly inside the actual deployment harnesses they'll run in, things like Codex, OpenClaw, and ZeroClaw, rather than a simplified stand-in that gets swapped out later. That mismatch between training and deployment has been a quiet source of agent flakiness for a while. By recording a harness's own model calls as training data through a lightweight proxy, Orchard closes that gap without forcing researchers to rebuild the harness logic themselves.
Microsoft is releasing the environment, the training data, and the evaluation methods alongside the models, which is the part that should actually move the field. Small models doing this well isn't just a cost story, it's a signal that the environment and training recipe matter as much as raw scale.
My take
I've been saying for a while that agent benchmarks are mostly theater until someone open-sources the actual training environment, and Orchard is exactly that overdue move. A 3-billion-parameter model getting within shouting distance of frontier systems on SWE-bench should worry anyone betting their moat on sheer parameter count rather than better infrastructure. This is the kind of open release that actually shifts power away from the handful of labs hoarding proprietary sandboxes, and I'd rather see ten more Orchards than one more closed 400-billion-parameter flex.
Read more about this at: Microsoft Research
