Microsoft open-sources Orchard to cut AI agent training costs

Why it matters

Orchard makes a costly layer of agent training available as open-source infrastructure, giving startups and labs a self-hosted alternative to managed sandbox vendors.

Microsoft Research said in an August 3rd post on X that it has released Orchard, an MIT-licensed framework for training and evaluating AI agents across coding, browser navigation and personal-assistant tasks. The project targets a costly problem behind agent development: researchers repeatedly building sandbox infrastructure, data pipelines and evaluation systems for each model and use case.

https://x.com/msftresearch/status/2084309571351253460?s=46

Orchard comes from a group led by Baolin Peng, a principal research manager whose work focuses on LLM-powered agents. Peng earned his computer science Ph.D. from the Chinese University of Hong Kong. The other lead authors named in Microsoft Research's launch post are Wenlin Yao, Qianhui Wu, Hao Cheng and Jianfeng Gao, with additional collaborators from Columbia University and the University of Illinois Urbana-Champaign.

The release turns agent infrastructure into a reusable service instead of embedding it inside a particular training stack. Orchard's core component, Orchard Env, runs on Kubernetes and exposes sandbox creation, command execution, file access and network controls through REST and Python interfaces. Researchers can connect different agent harnesses, trainers and inference systems without rebuilding the environment layer for each experiment.

That design addresses an increasingly expensive part of agent development. A software-engineering rollout may require cloning a repository, installing dependencies, editing files and running tests inside an isolated container. Reinforcement-learning experiments can repeat that process across thousands of concurrent environments. Managed services simplify provisioning, but bind researchers to an outside control plane and its pricing.

A self-hosted alternative to managed sandboxes

The Orchard paper compares the environment service with managed platforms including E2B, Daytona and Modal, as well as integrated research systems such as MegaFlow. Microsoft's differentiator is control: Orchard Env can be self-hosted on a standard Kubernetes cluster and used independently of the agent harness or training framework running above it.

Microsoft's cost figures are estimates rather than customer bills. Using prices collected in April 2026, the researchers modeled a workload of 128 sandboxes, each with two virtual CPUs and 8 GiB of memory, running for 240 hours. Orchard Env was estimated to cost $3,362 on on-demand instances or $673 using spot capacity. The same paper estimated costs of $7,078 for E2B or Daytona and $10,305 for Modal under the modeled workload.

Those savings assume a team can operate Kubernetes, configure network policies and manage spot capacity. Orchard shifts spending away from a managed sandbox provider, but it does not remove the operational work. For research labs and agent startups already running clusters, that trade can be attractive because the same infrastructure can handle data generation, reinforcement-learning rollouts and final evaluations.

The GitHub repository includes the Orchard Env code under an MIT license. Its interfaces cover sandbox lifecycle management, command execution, file I/O, patch application and network isolation. Microsoft Research says the service launched 1,000 sandboxes in parallel with a 100% success rate during its tests, completing the process in 26 seconds.

Three recipes test the same infrastructure

Microsoft Research released three domain-specific recipes to show how the shared layer can support different agents. Orchard-SWE trains software-engineering agents, Orchard-GUI targets browser navigation, and Orchard-Claw covers productivity work involving tools such as email and calendars.

For Orchard-SWE, the researchers collected 107,185 multi-turn software-engineering trajectories across 2,788 repositories. The released dataset contains successful and unsuccessful attempts, giving researchers negative examples for reward modeling and failure analysis. The paper reports that Orchard-SWE reached 69.7% on SWE-bench Verified using a model with roughly 3 billion active parameters, rising to 73% when a separate value model reranked multiple candidate solutions.

Orchard-GUI used 400 distilled demonstrations and 2,200 open-ended training tasks. Microsoft Research reports an average success rate of 68.4% across WebVoyager, Online-Mind2Web and DeepShop. Orchard-Claw was trained on 200 synthetic productivity tasks and reached a 59.6% pass rate when allowed three attempts, increasing to 73.9% when paired with the ZeroClaw harness.

The benchmark results are reported by Orchard's authors and span different test sets, harnesses and inference setups, so the percentages are not a single ranking of the three agents. Their common result supports Microsoft's central claim: smaller open-weight models can improve materially when researchers reuse rollouts, train inside realistic harnesses and preserve prior experiments instead of discarding them.

Microsoft packages a project that began in May

The August 3rd launch packages work that had already begun appearing publicly. The first Orchard paper was submitted on May 14th, while the current third revision was posted on July 30th. The repository says Orchard Env and the trajectory datasets arrived in May, followed by OpenWebRL work in June and OpenForge RL in July.

That timeline matters because Microsoft Research is releasing a growing infrastructure layer, rather than a single benchmark model. The software, datasets and recipes give outside teams a starting point for running agent training on their own clusters. For vendors selling managed agent sandboxes, Orchard also creates an open reference implementation against which pricing, latency and portability claims can be tested.