Field note
Orchard is not primarily another agent orchestrator. Its center is Orchard Env, a lightweight Kubernetes-native service for sandbox lifecycle management across task domains, agent harnesses, and training stages. The environment boundary becomes reusable infrastructure for rollout collection, reinforcement learning, evaluation, and inference-time experiments.
On top of that substrate, the paper presents software-engineering, GUI, and personal-assistant recipes. Orchard-SWE combines sparse and dense reinforcement signals, on-policy distillation, rubric-based process rewards, and historical-experience distillation. The reported benchmark results are research claims, but the architecture matters independently of the scores.
A shared environment layer reduces one of agent research's quiet failures: training in a bespoke harness and deploying into a different runtime with different tools, timeouts, filesystem behavior, and reset semantics. Kubernetes is useful here as a control plane for reproducible sandbox creation, teardown, resource policy, and experiment lineage, not as a marketing label.