Recoleta
GitHub
Source note

Orchard: An Open-Source Agentic Modeling Framework

Authors
Baolin Peng, Wenlin Yao, Qianhui Wu, Hao Cheng, Xiao Yu, Rui Yang, et al.
Source
arXiv
Published
Collection
Software Intelligence
Agentic LLM TrainingCode IntelligenceSoftware AgentsReinforcement LearningKubernetes Sandboxes
Open arXivMarkdown source

Summary

Orchard is an open-source agent-training system built around a thin Kubernetes environment service. It reports strong results for software-engineering, browser-use, and personal-assistant agents using reusable sandbox infrastructure plus SFT and RL recipes.

Problem

  • Agentic LLM training needs many isolated environments for code execution, web tasks, and tool use; closed or tightly coupled sandbox systems make results harder to reproduce and reuse.
  • Software-engineering agents need large trajectory sets and better learning from failed attempts, because SWE-bench-style rewards are sparse.
  • Open research needs lower sandbox cost and higher rollout throughput so smaller teams can train and evaluate agents at scale.

Approach

  • Orchard Env runs as a Kubernetes-native sandbox service with a REST API for sandbox creation, command execution, file I/O, network policy, and cleanup.
  • It injects a lightweight in-pod execution agent into user Docker images at runtime, so task images do not need rebuilds.
  • Execution requests go directly to sandbox Pod IPs, which avoids Kubernetes exec and WebSocket overhead on the hot path.
  • Orchard-SWE distills 107K software-engineering trajectories, keeps useful parts of unresolved trajectories with credit-assignment SFT, and uses Balanced Adaptive Rollout for sparse-reward RL.
  • Orchard-GUI and Orchard-Claw reuse the same environment layer for browser navigation and productivity-tool tasks with SFT plus RL.

Results

  • Orchard-SWE with Qwen3-30B-A3B-Thinking reaches 64.3% on SWE-bench Verified after SFT and 67.5% after SFT+RL under mini-swe-agent, claimed as a new open-source best among comparable-size models.
  • Orchard-GUI trains a 4B vision-language agent with 0.4K distilled trajectories and 2.2K open-ended tasks, reaching 74.1% on WebVoyager, 67.0% on Online-Mind2Web, 64.0% on DeepShop, and 68.4% average.
  • Orchard-Claw trains on 0.2K synthetic tasks and reaches 59.6% pass@3 on Claw-Eval; with the ZeroClaw harness it reaches 73.9% pass@3.
  • Orchard Env reports 0.280 s average command latency, compared with 0.747 s for E2B and 2.046 s for Modal in the cited benchmark setup.
  • In a 1,000-sandbox stress test, Orchard Env reports 100% success, 26 s end-to-end time, 11.75 s average create time, and 0.28 s average exec latency.
  • Estimated cost for 128 parallel 2-vCPU, 8-GiB sandboxes over 240 hours is $3,362 on-demand and $673 with spot instances, compared with $7,078 for Daytona or E2B.

Link