Microsoft Research Opens Orchard Framework for Scalable Agentic AI

What Microsoft Research introduced

Microsoft Research published Orchard on August 3, 2026 as an open framework for scalable, cost-conscious agentic AI research. Its central component, Orchard Env, separates reusable task environments from a particular model or agent harness so researchers can train and evaluate different agents against the same infrastructure.

The goal is to reduce duplicated engineering across software-engineering, browser-navigation and personal-assistant research. Rather than rebuilding environment setup, execution, reset, logging and evaluation for every new agent, Orchard exposes those capabilities as reusable services.

Training inside real agent harnesses

A notable design choice is that Orchard can train agents directly inside deployment-oriented harnesses rather than only in simplified benchmark wrappers. Microsoft says its infrastructure can work with systems such as Codex, OpenClaw and ZeroClaw, allowing the same environment and trajectory pipeline to be reused across different agent implementations.

That matters because agent performance often depends heavily on the harness: context management, tool interfaces, retry behavior, environment state and action policies can change results even when the underlying model remains the same. Training closer to the intended runtime can reduce the gap between benchmark conditions and deployed behavior.

Orchard-SWE, Orchard-GUI and Orchard-Claw

Microsoft describes three example stacks built on the framework. Orchard-SWE targets software engineering and includes a large trajectory dataset; Orchard-GUI focuses on web navigation; and Orchard-Claw explores personal-assistant-style agent tasks. The published blog reports benchmark results for these systems, but those figures should be read within Microsoft's implementation and evaluation setup rather than treated as model-independent guarantees.

The broader contribution is the shared infrastructure beneath those examples: reusable environment management, data collection and evaluation workflows that can support multiple task domains.

Why an environment layer matters

As agent systems become more complex, reproducibility increasingly depends on more than model weights. A benchmark run may rely on a particular browser state, repository version, operating-system image, network policy, tool set or hidden evaluator. Orchard's environment-service approach is designed to make those dependencies explicit and reusable.

For research teams, this can make it easier to compare models and training strategies under consistent task conditions. For open-source agent developers, it can also reduce the cost of collecting trajectories and experimenting with reinforcement-learning or imitation-learning pipelines.

Release status

Orchard is an open research framework rather than a new Microsoft foundation model. Its value is in agent training, evaluation and environment infrastructure, especially for teams that want to reuse the same task environments across different models and harnesses.