Orchard: An open framework for scalable agentic AI

Microsoft Research released Orchard, an open-source framework centered on a Kubernetes service that creates isolated environments for training and evaluating agents across coding, browser and personal-assistant tasks. It can train models inside deployment harnesses including Codex, OpenClaw and ZeroClaw.

Orchard-SWE reaches 69.7% on SWE-bench Verified with roughly three billion active parameters and 73% after value-model reranking. Orchard-GUI averages 68.4% across three browser benchmarks after training on 400 demonstrations and 2,200 tasks. Orchard-Claw reaches 59.6% on productivity tasks with three attempts and 73.9% in the stronger ZeroClaw harness.

Microsoft is releasing training data, evaluation methods and three task-specific recipes alongside the framework.

Agent capability is partly a property of the harness: tools, state management, retries and execution context shape the result. Training a bare model before placing it in a complex runtime creates a mismatch that ordinary model benchmarks overlook.

Signal Desk's read: separating the environment as reusable infrastructure is Orchard’s strongest contribution. It lets researchers compare models and learning methods against a more consistent execution layer while training in the same harness used at deployment. The cost is operational complexity: Kubernetes isolation and network controls demand serious administration.

This is a vendor-authored Microsoft Research post about Microsoft’s framework, so benchmark and state-of-the-art claims are vendor-reported. Open code and data should enable scrutiny, but benchmark success still does not demonstrate safety on unrestricted email, browser or coding work.

The compact-model results also argue against brute-force scale as the only path. Better task environments, learning from useful portions of failed attempts, richer rewards and candidate reranking can extract more from smaller models, although reranking adds attempts and cost.

Agent research often relies on proprietary sandboxes and task-specific infrastructure. Real agents operate through stateful harnesses handling tools and multiple processes, while many training systems rely on simplified substitutes.

Orchard’s reproducibility promise will depend on how faithfully its released environments capture the messy dependencies of real deployments. A common container interface is valuable, but it is not the same thing as common task difficulty or comparable agent autonomy.