Orchard is an open-source framework built around Orchard Env, a lightweight Kubernetes-native environment service that provides reusable, isolated components for training and evaluating agentic AI across multiple task domains. The framework aims to make agentic AI research more scalable and cost-effective by allowing researchers to reuse environments, data pipelines, and evaluation workflows across tasks and harnesses.
Why Orchard?
State-of-the-art agentic systems often rely on proprietary infrastructure—custom sandboxes, closed training pipelines, and private datasets—which many researchers cannot access or reproduce. Orchard addresses this gap by separating the runtime environment into a standalone, reusable service rather than embedding it within a single training framework.
Orchard Env — an environment layer that scales
Orchard Env runs on Kubernetes and can create, manage, and remove thousands of isolated components in parallel. It is designed to support diverse task types (coding, web browsing, tool use) and different agent systems, covering stages of the training and evaluation pipeline such as data distillation and reinforcement learning rollouts.
A key capability is training agents directly inside real deployment harnesses. Many capable agents operate through stateful, multi-process harnesses—examples include Claude Code, Codex, OpenClaw, and ZeroClaw—that coordinate multi-turn reasoning, tool use, and external integrations. Orchard provides a lightweight proxy that records the harness’s model calls as training data while each rollout runs in its own container, enabling end-to-end training in the same harness used at deployment and reducing mismatch between training and production.
Three domain-specific recipes and their performance
The project releases three domain-specific training recipes—Orchard-SWE (software engineering), Orchard-GUI (browser navigation), and Orchard-Claw (personal assistant)—along with training data and evaluation methods used to build them.
Orchard-SWE — software engineering agents
Software engineering tasks demand multi-step reasoning over real codebases, tool usage, and error recovery. Orchard-SWE is built on the Mini-SWE-Agent framework and evaluated on the SWE-bench Verified benchmark, which measures a model’s ability to navigate, diagnose, and repair real-world repositories.
For training, the team distilled 107,000 agent interactions from two advanced open-weight models (MiniMax-M2.5 and Qwen3.5-397B) spanning a wide range of GitHub Issues. The supervised fine-tuning used credit-assignment: rather than discarding partially successful attempts, the process learns from productive segments of those attempts to expand usable training data.
Reinforcement learning follows, but feedback is sparse—typically only whether the final patch passed hidden tests. Orchard starts with Balanced Adaptive Rollout to make the most of infrequent success signals, then adds two dense-reward techniques for richer guidance: on-policy distillation, where a stronger teacher model scores decisions step by step, and a process reward model, where an AI judge rewards sound problem-solving behavior (writing reproducing tests, verifying fixes, checking regressions) independent of final pass/fail outcomes.
A value model is trained on previous rollouts to rerank candidate solutions. Instead of discarding practice trajectories, data from 20 prior experiments train a compact value model (~4 billion parameters) that recognizes high-quality solutions and reranks multiple candidate answers at inference time.
Together, these techniques raise Orchard-SWE from a 61.4% baseline on SWE-bench Verified to 69.1% with Balanced Adaptive Rollout and 69.7% with dense-reward methods—establishing a new state of the art among open-source models of comparable size (roughly 3 billion active parameters). With value-model reranking the score reaches 73.0%, approaching frontier systems more than ten times larger.
Orchard-GUI — a lightweight browser agent
Web navigation requires understanding visual layouts, interacting with dynamic interfaces, and solving open-ended tasks given only natural-language instructions. Orchard-GUI trains a 4-billion-parameter vision-language model as a browser agent using a relatively small supervised dataset: 400 distilled demonstrations plus 2,200 open-ended training tasks.
Despite the limited supervision, the model achieves strong results across several web-navigation benchmarks: 74.1% on WebVoyager, 67.0% on Online-Mind2Web, and 64.0% on DeepShop, for an average success rate of 68.4%. These results place Orchard-GUI among the strongest open-source GUI agents while remaining competitive with larger proprietary systems.
Orchard-Claw — personal assistant agents
Many impactful agentic applications involve everyday productivity tasks such as reading and drafting emails, managing calendars, searching for information, and coordinating across tools. Orchard-Claw focuses on these personal-assistant tasks by training an agent on only 200 synthetic tasks. Evaluated on Claw-Eval, a benchmark covering realistic productivity workflows, it completes 59.6% of tasks given up to three attempts; this rises to 73.9% when paired with the stronger ZeroClaw agent system.
Because Orchard trains agents directly inside real deployment harnesses, Orchard-Claw is trained across several harnesses—including ReACT, ZeroClaw, OpenClaw, and Codex—rather than a single simplified loop. Training in real harnesses significantly improves reliability: under the Codex harness, for example, success increases from 18.6% for the untrained model to 51.5% after Orchard training.
Implications and next steps
Orchard’s results underscore that the environment layer matters. By making the underlying infrastructure open, lightweight, and reusable, Orchard lowers the cost and barrier to entry for agentic AI research. Teams can generate training data, run reinforcement learning rollouts, and evaluate models without rebuilding infrastructure each time or relying on proprietary cloud services.
The project also advocates reusing training experience as a path toward cumulative agent learning: rather than discarding trajectories after a run, they can be treated as persistent assets (for example, distilled into reusable value models), allowing successive agent generations to inherit and extend prior knowledge.
The data efficiency shown by Orchard-GUI suggests larger-scale web agents could be trained without enormous amounts of hand-crafted training data. By releasing the full Orchard stack (environment service, training pipelines, and datasets), the authors aim to help the broader research community build and study capable open agentic systems more rapidly.
Acknowledgements
The Orchard authors thank teams at Microsoft Research and collaborating institutions, as well as the open-source community whose benchmarks and tools enabled the research.
