In a post published by Microsoft Research, Orchard was introduced as an open-source framework designed for scalable and cost-effective agentic AI research. The project is built around Orchard Env, a reusable environment service designed to train and evaluate agents across a variety of task domains. The same underlying infrastructure supports agents designed for software engineering, web navigation, and personal assistants, allowing them to be trained directly inside real deployment harnesses such as Codex, OpenClaw, and ZeroClaw. This approach enables researchers to reuse work environments, data pipelines, and evaluation workflows across different tasks.
Introduction: Solving the Bottleneck in AI Agent Research
Artificial intelligence technology is rapidly advancing beyond static question-answering toward autonomous agents capable of planning, reasoning, and acting within complex, multi-step environments. These systems are capable of fixing bugs in complex source code, navigating the web on a user's behalf, and managing workflows that include calendars and email.
Despite the excitement surrounding these capabilities, the research community faces a persistent bottleneck. Developing advanced agentic systems often requires proprietary infrastructure, including custom sandbox environments, closed training pipelines, and proprietary datasets that most researchers and professionals cannot access or reproduce.
To address this gap, researchers have introduced Orchard—an open-source platform for scalable agent modeling. At the core of the project is Orchard Env, a lightweight Kubernetes-based environment that provides isolated, reusable components for running and building agents at scale—from collecting training data to reinforcement learning rollouts and evaluation.
The Architecture of Orchard Env: Separating the Environment Layer
The central concept behind Orchard is that the runtime environment should operate as a standalone, reusable service rather than infrastructure embedded within a specific training framework. Orchard Env's Kubernetes foundation enables it to create, manage, and remove thousands of isolated components in parallel.
The system is designed to work across diverse tasks such as code writing, web browsing, and tool use. It is also optimized to work with different agent systems at various stages of the training and evaluation processes, including data distillation and reinforcement learning rollouts. This flexibility makes Orchard practical at a research scale, allowing teams to introduce new benchmarks, agent systems, or training algorithms without rebuilding the underlying infrastructure from scratch.
Training Agents Directly Inside Real Deployment Harnesses
One of the unique features of Orchard is its ability to train agents directly inside real deployment harnesses. Today, the most capable agents rarely operate as a standalone, clean model; instead, they run through sophisticated harnesses—such as Claude Code, Codex, and OpenClaw—that manage multi-step reasoning, tool use, and connections to external systems.
Open training tools are typically unable to handle these stateful, multi-process harnesses, forcing researchers to train models on simplified stand-ins and then deploy them in real environments, which creates a performance mismatch. Orchard closes this gap: a lightweight proxy records the harness's own model calls as training data, while each rollout runs in its own container. This makes it possible to train an agent end-to-end directly inside the deployment harness in which it is intended to run—such as OpenClaw, Codex, ZeroClaw, and others—and even across multiple harnesses simultaneously.
Orchard-SWE: A Breakthrough in Open-Source Software Engineering Agents
Software engineering is one of the most demanding domains for autonomous agents, as it requires multi-step reasoning over real source code, tool use, and the ability to recover from errors. Orchard-SWE is the researchers' workflow for training in this domain. It is built on the Mini-SWE-Agent framework, which is designed to autonomously solve software engineering tasks, and is evaluated on the widely used SWE-bench Verified benchmark, which tests a model's ability to navigate, diagnose, and repair real-world source code.
To train the system, researchers distilled 107,000 agent interactions from two advanced open-weight models (MiniMax-M2.5 and Qwen3.5-397B), covering a broad range of GitHub issues. The training process utilizes credit-assignment supervised fine-tuning: instead of completely discarding attempts where the agent failed to fully resolve the issue, the system learns from the productive portions of those partial attempts, thereby expanding the amount of useful training data available to the model.
In the next phase, reinforcement learning is applied, but the feedback is sparse—an agent typically learns only whether its final patch passed or failed hidden tests. Researchers begin with Balanced Adaptive Rollout, designed to make the most of those rare success signals, and then add two "dense reward" techniques for richer guidance: on-policy distillation, where a stronger teacher model scores the agent's decisions step-by-step, and a process reward model, where an AI judge rewards a sound problem-solving process—such as writing tests that reproduce the bug, verifying the fix, and checking that existing behavior still works—regardless of whether the final tests passed successfully.
Finally, a value model is trained on past rollouts to rerank candidate solutions. Reinforcement learning generates many practice trajectories that are typically discarded; instead, trajectories from 20 prior experiments train a compact, 4-billion-parameter value model that identifies high-quality solutions. At problem-solving time, it ranks multiple candidate answers and selects the best one.
Together, these techniques elevate Orchard-SWE from a baseline of 61.4% on SWE-bench Verified to 69.1% with Balanced Adaptive Rollout, and to 69.7% with the dense reward techniques. This establishes a new state-of-the-art performance benchmark among open-weight models of comparable size (roughly 3 billion active parameters, the 35B-A3B model). The score rises to 73% when combined with reranking via the value model, approaching the performance of frontier systems that are 10 times larger or more.
Orchard-GUI: A Lightweight Browser Agent for Real-World Web Tasks
Web navigation presents a different set of challenges. Agents must interpret visual layouts, interact with dynamic interfaces, and complete open-ended tasks described only in natural language.
Orchard-GUI trains a 4-billion-parameter vision-language model as a browser agent using a relatively small amount of supervision: 400 distilled demonstrations combined with 2,200 open-ended training tasks. Despite this limited training data, the model achieves strong results across several web navigation benchmarks: 74.1% on WebVoyager, 67.0% on Online-Mind2Web, and 64.0% on DeepShop, averaging 68.4%.
These results position Orchard-GUI among the strongest open-source web agents today, while remaining competitive against larger proprietary models from OpenAI and Google. These findings demonstrate that with the right training approach and environment, small open models can deliver excellent performance on real-world web tasks.
Orchard-Claw: Personal Assistant Agents for Everyday Productivity
Many of the most impactful agent applications involve everyday productivity tasks, including reading and writing emails, managing calendars, searching for information, and coordinating across different tools. Orchard-Claw focuses on personal assistant tasks by training an agent on just 200 synthetic tasks.
Evaluated on Claw-Eval, a benchmark covering realistic productivity workflows, the agent successfully completed 59.6% of tasks when given up to three attempts. The success rate rose to 73.9% when paired with the stronger ZeroClaw agent system.
Because Orchard is capable of training agents directly inside real deployment harnesses, Orchard-Claw was trained across several of these harnesses—including ReACT, ZeroClaw, OpenClaw, and Codex—rather than in a single simplified loop. Training inside these real harnesses significantly improved the agent's reliability; under the Codex harness, for example, its success rate rose from 18.6% for the untrained model to 51.5% after training with Orchard.
Implications for the Future and Accumulating Experience in Agent Research
The results of the Orchard project reinforce a broader insight: the environment layer is of critical importance. By making the underlying infrastructure open, lightweight, and reusable, Orchard reduces the cost of researching AI agents. Teams no longer need to build custom isolated environments from scratch or rely on proprietary cloud services. The same Orchard Env can be used to generate training data, run reinforcement learning rollouts, and evaluate final models without needing to rebuild the system each time.
Looking ahead, researchers identify the reuse of training experience as a promising path toward cumulative agent learning. Instead of discarding trajectories at the end of a training run, the system treats them as persistent assets—for example, by distilling them into reusable value models. This allows agent experience to accumulate over time, enabling each new generation of agents to inherit and extend the knowledge acquired by its predecessors, rather than starting from scratch. The data efficiency demonstrated in Orchard-GUI suggests that larger-scale web agents could be trained without requiring vast amounts of manually created training data.