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

Microsoft Orchard: Decoupling the Environment Layer from Agent Training Stacks

Forum topic · 小凯 · 2026-08-05

Summary

Microsoft Research open-sourced Orchard, a Kubernetes-native environment service (Orchard Env) plus Python SDK that spawns thousands of isolated containers and exposes sandbox lifecycle, command execution, file I/O, network control, and agent integration over HTTP. By treating environments as a reusable neutral service rather than a per-task component, the same infrastructure powers SWE, GUI, and personal-assistant (Claw) agent training. Cost is roughly 0.47x Daytona per 2-vCPU/8-GiB sandbox, or 0.10x on spot instances. Benchmarks include Orchard-SWE (Qwen3.5-35B-A3B reaching 73% on SWE-bench Verified), Orchard-GUI (4B VLM averaging 68.4% across WebVoyager, Online-Mind2Web, DeepShop), and Orchard-Claw (30B-A3B at 51.5% under the Codex harness). The companion OpenForge RL bridges the train-deploy gap by replaying real deployment-harness traces into RL frameworks, enabling rollouts inside production harnesses such as OSWorld-Verified and QwenClawBench.

What Orchard Is

Orchard is not an agent framework. It is the *environment layer* for agent training — a reusable service that solves the engineering pain of running tens of thousands of parallel agent experiments without standing up an isolated sandbox for every single run.

Core Architecture: Orchard Env

  • Kubernetes-native environment service + Python SDK that provisions thousands of isolated containers on demand.
  • Exposes generic capabilities over HTTP: sandbox lifecycle, command execution, file I/O, network control, agent integration.
  • Harness-neutral: the same Env supports SFT trajectory distillation, RL rollouts, and evaluation task switching — no rebuild needed when the task changes.
  • In the arXiv comparison table, Orchard Env is the only open-source entry that fits the "narrow, independent, reusable" definition for a thin environment service.
  • Cost: ~0.47x Daytona per 2-vCPU/8-GiB sandbox; ~0.10x with spot instances.
  • Three Training Recipes Sharing One Stack

    Orchard-SWE

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

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

  • 30B-A3B personal-assistant agent, starting from 200 synthetic tasks.
  • pass@3 on Claw-Eval: 59.6%; with the ZeroClaw harness: 73.9%.
  • Training runs inside real deployment harnesses (ReACT, ZeroClaw, OpenClaw, Codex), not a "simplified loop".
  • Under the Codex harness: 18.6% → 51.5%.
  • OpenForge RL: Bridging the Train-Deploy Gap

  • Updated in July. Traditional RL stacks train on simplified harnesses; switching harnesses at deployment collapses performance.
  • OpenForge adds a lightweight proxy that records calls from the real deployment harness and reconstructs RL samples for any RL codebase (e.g., veRL).
  • Orchard Env spins up matching containers for rollouts in parallel.
  • Results:
  • OpenForge-GUI 8B: OSWorld-Verified 37.7, WebVoyager 72.3.
  • OpenForge-Claw 30B-A3B: QwenClawBench 33.7, MCPAtlas 28.1.
  • Why It Matters

    The signal isn't in the scores — it's in the taxonomy. Orchard decouples "environment" from the training stack as a reusable service. That single abstraction lets one infrastructure layer serve SWE, GUI, and Claw domains simultaneously.

    The next phase of agent training is unlikely to be "bigger models + more trajectories". It is more likely to be cheaper, standardized environment pools + cross-harness RL. For teams building agent training frameworks — whether on LangGraph, ReAct, or in-house runtimes — the design lesson is clear: treat environments as services, not components, and make training a swappable backend.

    Resources

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

Tags

#microsoft-orchard#agent-training#environment-service#kubernetes#reinforcement-learning#openforge#swe-bench#gui-agent

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