Microsoft releases Orchard, an open-source framework that trains efficient AI agents for code, web, and productivity tasks

Microsoft Research has unveiled Orchard (opens in new tab), an open-source platform designed to streamline the development and training of autonomous AI agents across diverse real-world tasks—from fixing software bugs to navigating websites and managing email workflows.

At the heart of the system sits Orchard Env, a lightweight Kubernetes-based runtime that functions as a reusable service layer. Rather than embedding specialized infrastructure into individual training frameworks, this approach allows researchers to spin up thousands of isolated agent environments in parallel, share training data pipelines, and reuse evaluation methods across fundamentally different agent types.

Solving the infrastructure bottleneck#

The core problem Orchard tackles is a familiar one in AI research: building state-of-the-art agents typically demands proprietary sandboxes, closed training pipelines, and private datasets that remain inaccessible to most researchers. This creates a barrier to reproducibility and limits innovation outside well-funded labs.

Orchard flips this model by making the underlying infrastructure public and flexible. The same environment service handles software-engineering agents, web-browsing systems, and personal-assistant bots without modification. Researchers can introduce new benchmarks or algorithms without rebuilding the platform from scratch—a significant time and cost saving.

Another notable feature: Orchard allows agents to train directly within their actual deployment harnesses—systems like Claude Code, Codex, and OpenClaw—rather than in simplified proxies. Traditionally, researchers train on a stripped-down version and hope the model transfers to the real system when deployed. Orchard captures the harness's own model calls as training data while each trial runs in its own containerized sandbox, eliminating that mismatch.

Three domain-specific agents demonstrate capability#

Microsoft is releasing three pre-built agent systems built on Orchard:

• Orchard-SWE (software engineering): Trained on 107,000 interactions distilled from advanced open models, this agent reaches 67.5% on SWE-Bench Verified—73% with a value-reranking layer. That performance rivals frontier systems 10–30 times larger, using only ~3 billion active parameters. The training pipeline incorporates credit-assignment fine-tuning on partial successes, reinforcement learning with dense reward signals from teacher models and process monitors, and a compact value model that ranks candidate solutions.