Microsoft open-sources Orchard, a shared training environment for AI agents
Microsoft Research released Orchard, an open Kubernetes-native environment service for training and evaluating agents across coding, web, and personal-assistant tasks, plus three small model recipes that punch well above their parameter count.
What shipped
Microsoft Research released Orchard on August 3, an open-source framework for training and evaluating agentic AI systems, plus three domain-specific model recipes built on top of it: Orchard-SWE for software engineering, Orchard-GUI for web navigation, and Orchard-Claw for personal-assistant tasks. The team also released the training data and evaluation code behind all three.
The problem Orchard targets is mundane but real. Training an agent, as opposed to a static model, means running it inside an environment: a sandbox where it can execute code, click through a browser, or call a calendar API, fail, retry, and generate a reward signal. Every lab that does this seriously builds this environment layer internally, and it rarely gets published. Open and academic teams end up rebuilding a thinner version of the same infrastructure for each project, which is a large part of why open agent research tends to lag proprietary systems on anything past narrow benchmarks.
One environment, three domains
At the center of the release is Orchard Env, a Kubernetes-native service that spins up and tears down isolated sandboxes at scale — for code execution, browser sessions, or tool calls — behind a single REST API regardless of task type. The design choice worth noting: Orchard Env isn't bundled into a specific training framework. It's a standalone service a training loop calls into, so the same infrastructure handles data distillation, reinforcement-learning rollouts, and evaluation without being rewritten per project.
The more interesting piece is how Orchard handles the gap between training and deployment. Production agents run inside harnesses — Claude Code, Codex, OpenClaw — that manage multi-turn state, tool routing, and context across a session. Most open RL tooling trains against a simplified stand-in for that harness and then deploys into the real one, introducing a mismatch between what the agent learned and how it's actually used. Orchard's fix is a lightweight proxy that sits in front of the harness's own model calls and records them as training data, while each rollout runs in its own container. That lets a team train an agent directly inside the harness it will ship in, including across several harnesses at once — which is exactly what Orchard-Claw does.
The three recipes
Orchard-SWE pairs credit-assignment supervised fine-tuning — extracting useful signal from partial, failed repair attempts instead of discarding them — with reinforcement learning against SWE-bench Verified's sparse pass/fail signal, then adds a compact 4B-parameter value model trained on otherwise-discarded rollouts to rerank candidate patches at inference time. Result: 69.7% on SWE-bench Verified at roughly 3B active parameters, rising to 73% with reranking, approaching results from models more than 10x larger.
Orchard-GUI is a 4B vision-language browser agent trained on a comparatively small 2,600 examples, averaging 68.4% across WebVoyager, Online-Mind2Web, and DeepShop. Orchard-Claw, trained on just 200 synthetic tasks, hits 59.6% on Claw-Eval (73.9% paired with a stronger agent system on top) — and, the number I find most telling, goes from an 18.6% baseline to 51.5% under the Codex harness purely from being trained inside that harness rather than a stand-in for it.
Where I land
This is useful infrastructure, and the small-model results are genuinely good — a 3B-active-parameter model landing within striking distance of systems an order of magnitude larger is not nothing. But it's worth being precise about what's being claimed. The post doesn't benchmark Orchard Env itself against the agent-RL tooling that already exists in the open-source ecosystem — no comparison on throughput, scale, or ergonomics against the environment layers other groups have built for agent training. Without that, I can't tell whether Orchard is meaningfully better infrastructure or just infrastructure Microsoft happened to publish.
The harness-training proxy is the one idea here I'd actually go build on. Training inside the deployment harness instead of a stand-in is a real fix for a real mismatch, and the Codex numbers — 18.6% to 51.5% — show it matters in practice, not just in theory. The rest reads as a solid, well-documented contribution to a community that's been short on shared tooling: worth watching for adoption, not yet a result that changes what's achievable.