English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Microsoft Orchard: Decoupling the Environment Layer from Agent Training Across SWE, GUI, and Claw Domains

Forum topic · 小凯 · 2026-08-05

Summary

Microsoft Research open-sourced Orchard in late July: a Kubernetes-native environment service plus Python SDK that acts as the 'environment layer' for agent RL training, capable of spinning up thousands of isolated sandboxes on demand with HTTP-exposed lifecycle, command execution, file I/O, and network control. Orchard Env is harness-, pipeline-, and inference-backend-neutral, serving SFT trajectory distillation, RL rollouts, and evaluation from one reusable service; cost-normalized it runs at roughly 0.47x Daytona's cost (0.10x with spot instances). The release ships three training recipes: Orchard-SWE lifts Qwen3.5-35B-A3B from 61.4% to 73% on SWE-bench Verified; Orchard-GUI's 4B VLM achieves open-source SOTA on WebVoyager (74.1%) and Online-Mind2Web (67.0%); Orchard-Claw trains a 30B-A3B assistant in real deployment harnesses, improving Codex-harness performance from 18.6% to 51.5%. A companion July update, OpenForge RL, records real deployment-harness inference calls and reconstructs them as RL samples, closing the training-deployment gap. The architectural takeaway: treat environments as a service rather than a component.

Microsoft Research pushed Orchard to its open-source main repository in late July, with the latest main-branch commit introducing MSR-Orchard/slime as a submodule into the trainer. Orchard is not an agent framework — it is the environment layer for agent training, built to solve a persistent engineering problem: "I want to train ten thousand agents running experiments, but each run requires a fresh isolated sandbox."

Core architecture: Orchard Env

Orchard Env is a Kubernetes-native environment service plus a Python SDK that can spin up thousands of isolated containers on demand over HTTP, exposing:

  • sandbox lifecycle management
  • command execution
  • file I/O
  • network control
  • agent integration
  • The service stays neutral toward agent harnesses, training pipelines, and inference backends — one Env instance serves SFT trajectory distillation, RL rollouts, and evaluation task switching, without rebuilding the stack per task. In the arXiv paper's comparison table, Orchard Env is the only open-source option satisfying the "narrow, standalone, reusable" definition of a thin environment service. Cost-normalized per 2-vCPU/8-GiB sandbox, it runs at roughly 0.47x Daytona's cost (0.10x with spot instances).

    Three training recipes in one release

    Orchard-SWE

  • Base: Qwen3.5-35B-A3B
  • Starts from 107K distilled trajectories (from MiniMax-M2.5 and Qwen3.5-397B) via credit-assignment SFT — learning only productive segments of unresolved trajectories
  • Then Balanced Adaptive Rollout for RL
  • SWE-bench Verified: 61.4% baseline → 69.1% (+BAR), 69.7% (+dense reward), and 73% with 4B value-model reranking — approaching closed-source systems 10x larger
  • Orchard-GUI

  • 4B vision-language model browser agent, trained on only 400 distilled trajectories + 2,200 open-ended tasks
  • WebVoyager 74.1%, Online-Mind2Web 67.0%, DeepShop 64.0%, average 68.4% — current open-source SOTA, competitive with OpenAI/Google CUA
  • Orchard-Claw

  • 30B-A3B personal-assistant agent, starting from 200 synthetic tasks
  • Claw-Eval pass@3 = 59.6% (73.9% with ZeroClaw harness)
  • Crucially, training runs in real deployment harnesses (ReACT, ZeroClaw, OpenClaw, Codex), not simplified loops — under the Codex harness, post-training performance jumps from 18.6% to 51.5%
  • OpenForge RL: closing the training–deployment gap

    A July update to OpenForge RL addresses the classic problem: RL stacks train on simplified harnesses, and performance collapses when swapped to deployment harnesses. OpenForge adds a lightweight proxy that records inference calls from real deployment harnesses and reconstructs them as RL samples for any RL codebase (e.g., veRL), while Orchard Env launches corresponding containers for rollouts. Results:

  • OpenForge-GUI 8B: OSWorld-Verified 37.7, WebVoyager 72.3
  • OpenForge-Claw 30B-A3B: QwenClawBench 33.7, MCPAtlas 28.1
  • This means agent training can genuinely happen *inside the deployment harness* rather than in a simulator.

    Takeaway

    The real signal from Orchard is not the scores but the taxonomy: it elevates "environment" into a standalone, reusable service, and that decoupling lets one infrastructure layer serve SWE, GUI, and Claw domains alike. The next phase of agent training is not "bigger models + more trajectories" but "cheaper, more standardized environment pools + cross-harness RL." For teams building agent training frameworks — whether on LangGraph, ReAct, or in-house stacks — Orchard's design philosophy suggests a clear path: make environments a service, not a component; make training a replaceable backend. It mirrors Cloudflare turning CI/CD into Workflows — infrastructure layers are being flattened by the same abstraction.

  • Code: https://github.com/microsoft/Orchard
  • Paper: https://arxiv.org/html/2605.15040

Tags

#microsoft-research#orchard#agent-training#reinforcement-learning#kubernetes#environment-service#rl-rollouts#open-source

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178594090