Microsoft Open-Sources Orchard: Infrastructure for Training Agents in Real Environments
udmrzn · x · 2026-08-05
Microsoft has open-sourced Orchard, an extensible infrastructure for agent training, reinforcement learning, and evaluation.
- Core Value: Enables developers to train models directly within real agent harnesses like Codex, OpenClaw, and ZeroClaw, rather than relying on simplified experimental environments.
- Initial Offerings:
- Orchard-SWE: Achieves 69.7% on SWE-bench Verified (73% after re-ranking).
- Orchard-GUI: A 4B parameter model averaging 68.4% on web tasks.
- Orchard-Claw: Focused on email, calendar, and personal assistant tasks.
- Architecture: The core Orchard Env is built on Kubernetes, allowing for the batch creation of isolated environments while reusing data, training, and evaluation pipelines.
By combining small models with real-world environments and open data, it provides a critical foundation for open-source agent training.
Related event: Microsoft Open-Sources Orchard for AI Agent Training and Evaluation(2 posts)→
More from coding & agent
- A 9-Layer Stack Guide for Building Production AI Agents — MaryamMiradi · 2026-08-05
- Open-Source Channels SDK: Connect Any AI Agent to Slack and Teams — _jaydeepkarale · 2026-08-05
- "Verification Sharding": Distributing Output Validation Across Massive Agent Swarms — curious_vii · 2026-08-05
- Building Software on the Eve of ASI: Six Core Opportunities — pzakin · 2026-08-05
- Securely Running AI Agents on Kubernetes: Zero-Trust and Sandbox Practices — WirelessLife · 2026-08-05
- Claude Agent Manages $50K Portfolio, Beats S&P 500 and Attracts $27M Copy-Trading — emeka_boris · 2026-08-05