Morph1k visual note · 01Microsoft open-sourced the floor agents train on
Orchard turns sandboxes, datasets, and evaluation into a reusable Kubernetes service — and ships recipes for coding, browser, and assistant agents.
Primary reading
Microsoft Research Orchard announcementWhat happened
On August 3 Microsoft Research released Orchard, an MIT-licensed framework built around Orchard Env — a Kubernetes service for spinning up thousands of isolated sandboxes. It includes three training recipes (Orchard-SWE, Orchard-GUI, Orchard-Claw), public datasets, and evaluation methods. Orchard-SWE reaches 69.7% on SWE-bench Verified with about 3B active parameters, rising to 73% with value-model reranking.
Why it matters
Agent progress has been stuck behind proprietary sandboxes. A reusable environment layer lets teams train inside real harnesses — Codex, OpenClaw, ZeroClaw — instead of a simplified stand-in that never matches production.
For you
If you build or fine-tune agents, inspect Orchard before rebuilding sandbox plumbing. Start with the published SWE or GUI recipe if your stack runs on Kubernetes; treat the harness-in-the-loop path as the interesting part.
