Orchard: Revolutionizing Agentic AI with an Open-Source Framework

Orchard: An Open-Source Framework for Evolving Agentic AI

Orchard positions itself as an innovative open-source framework dedicated to research on evolving agentic artificial intelligence. At the heart of this system is Orchard Env, a reusable environment service that facilitates the training and evaluation of agents across various task domains. This unique infrastructure allows researchers to work with agents in diverse areas such as software development, web navigation, and personal assistants, while utilizing real deployment environments like Codex, OpenClaw, and ZeroClaw. With Orchard, researchers can reuse environments, data pipelines, and evaluation workflows, significantly simplifying the research process.

The same Orchard infrastructure is capable of supporting agents in diverse fields such as software development, web navigation, and personal assistance. This means that researchers can train these agents directly in real deployment environments, which is a significant advantage. For example, Orchard-SWE, Orchard-GUI, and Orchard-Claw demonstrate that even relatively small open-weight models can achieve impressive results on complex real-world tasks. Orchard-SWE, for instance, achieves a score of 69.7% on the SWE-bench Verified benchmark, using only about 3 billion active parameters. This brings it close to state-of-the-art systems that utilize much larger models. Furthermore, with the value model reranking, Orchard-SWE reaches 73.0%, making it even more competitive.

Data Sharing and Evaluation Methods

In addition to providing models and workflows, Orchard offers training data and evaluation methods. These resources are intended to help the research community build and study open agentic systems. Artificial intelligence is rapidly evolving, moving from simple question answering to autonomous agents capable of planning, reasoning, and acting in complex, multi-step environments. These systems can correct bugs in complex codebases, navigate the web on behalf of a user, and manage workflows involving calendars and emails.

Challenges and Solutions for Agentic AI

Despite the growing enthusiasm around the capabilities of agentic AI, the research community faces a major hurdle: the need for proprietary infrastructure. This often includes custom testbeds, closed training pipelines, and proprietary datasets that are inaccessible to most researchers and practitioners. Orchard aims to bridge this gap by offering an open-source framework for evolving agentic modeling. At the core of Orchard is Orchard Env, a lightweight environment based on Kubernetes, which provides reusable isolated components to execute and build agents at scale.

Unlike many existing frameworks, Orchard Env is designed to support different agent systems and task types without requiring modifications. The same service can support software development agents, web navigation agents, and personal assistance agents across various domains. To demonstrate this approach, three domain-specific training recipes have been published: Orchard-SWE, Orchard-GUI, and Orchard-Claw. The training data and evaluation methods used to construct them are also made available.

An Adaptable Environment Layer

The central idea behind Orchard is that the execution environment should be a standalone, reusable service rather than an infrastructure embedded in a specific training framework. Thanks to its Kubernetes foundation, Orchard can create, manage, and delete thousands of isolated components in parallel. The system is designed to operate across tasks such as programming, web navigation, and tool usage, as well as different agent systems and stages of the training and evaluation process.

Orchard also allows agents to be trained within any environment. Today's top-performing agents rarely operate as a bare model. They function through sophisticated environments—such as Claude Code, Codex, and OpenClaw—that manage multi-turn reasoning, tool usage, and connections to external systems. Open training tools generally cannot handle these multi-process and stateful environments, forcing researchers to train on a simplified model and then deploy in the real framework, creating a gap. Orchard fills this void: a lightweight proxy records the model calls from the environment itself as training data while each deployment runs in its own container, allowing an agent to be trained end-to-end directly in the environment with which it will be deployed—OpenClaw, Codex, ZeroClaw, or others—and across multiple environments.

Orchard-SWE: Pioneering Software Development Agents

Software development represents a particularly demanding environment for autonomous agents. Orchard-SWE is a training workflow specifically designed for this domain. It uses the Mini-SWE-Agent framework to autonomously solve software development tasks and is evaluated on the SWE-bench Verified benchmark. To train the system, 107,000 agent interactions were distilled from advanced models, covering a wide range of GitHub issues.

The training process of Orchard-SWE utilizes supervised credit assignment refinement, allowing the system to learn from productive partial attempts. Reinforcement learning is also employed, although its feedback is rare. Dense reward techniques, such as on-policy distillation and a process reward model, are added for richer guidance. Finally, a value model is trained on past deployments to reevaluate candidate solutions. Orchard-SWE improves from a baseline of 61.4% to 69.1% with the Balanced Adaptive Rollout and reaches 69.7% with dense reward techniques, ultimately achieving 73.0% with the value model reranking.

Orchard-GUI: A High-Performing Browser Agent

Web navigation presents unique challenges, requiring agents capable of interpreting visual layouts and interacting with dynamic interfaces. Orchard-GUI trains a 4 billion parameter vision-language model as a browser agent, using limited supervision. Despite this, the model achieves good results on several web navigation benchmarks: 74.1% on WebVoyager, 67.0% on Online-Mind2Web, and 64.0% on DeepShop, averaging 68.4%. These results place Orchard-GUI among the top-performing open-source web agents to date.

Orchard-Claw: Agents for Everyday Productivity

Orchard-Claw focuses on personal assistance tasks, training an agent on just 200 synthetic tasks. Evaluated on Claw-Eval, it successfully completes 59.6% of tasks when given up to three attempts. This figure increases to 73.9% when paired with the more powerful agent system ZeroClaw. Because Orchard can train agents directly within real deployment environments, Orchard-Claw is trained across several of them, including ReACT, ZeroClaw, OpenClaw, and Codex. Under the Codex environment, its success rate increases from 18.6% for the untrained model to 51.5% after Orchard training.

Implications for AI Research

The results from Orchard highlight the importance of the environment layer in agentic AI research. By making the underlying infrastructure open, lightweight, and reusable, Orchard reduces the cost of research and eliminates the need to build custom silos.