On 4 August 2026, Microsoft Research open-sourced Orchard, a framework that finally separates AI agent training from execution. It ships a Kubernetes-native environment service and three domain recipes — Orchard-SWE, Orchard-GUI, and Orchard-Claw — that together turn agent research from a bespoke infrastructure project into a reusable stack. For teams building mobile automation, AI agents, and orchestration graphs, like the work we ship at Halmob, it is the clearest signal yet that the missing layer of the agent stack has moved from labs to public GitHub.
Every serious agent team has quietly built the same thing twice: a sandbox that lets a model open files, call tools, run commands, and be graded against a real task. Each rebuild costs weeks and yields infrastructure that only the team who wrote it can use. Orchard collapses that work into a shared service and layers three trained agents on top so a research team, a product team, or a Halmob n8n automation squad can pick up where the last one left off.
The 30-Second Version
What Microsoft Orchard Actually Is
Orchard is a modeling framework, not a single model. The design starts from a plain observation: the environment an agent runs in should be a reusable service, not code that lives inside one training framework. Once the environment is a shared service, any training method — SFT, rejection sampling, GRPO, RPR — can borrow it, and the same trained agent can be executed inside more than one production harness.
The Microsoft Research announcement post frames the release as three pieces that fit together:
- Orchard Env: a lightweight Kubernetes-native environment service. Sandbox management, command execution, file access, network controls, a REST API, and agent integration are all exposed as a shared capability rather than one team's private harness.
- Three training recipes: Orchard-SWE, Orchard-GUI, and Orchard-Claw. Each is a full recipe — data, method, and a trained model — for a different agent domain.
- Trajectory datasets: the multi-turn rollouts used to train each recipe are released on Hugging Face, so a research or product team can retrain, mix, or extend them without rerunning the collection.
Agent frameworks in 2026 are converging on the same three-layer shape: a reusable environment, a training loop that does not own the environment, and a harness that runs the trained agent in production. Orchard is the first open release that ships all three at once.
Orchard Env: The Shared Runtime
Orchard Env is the piece that most teams will actually touch first. It is a Kubernetes-native service that takes an agent request and gives it a sandbox — a fresh working directory, a filesystem, network controls, and the ability to run commands. The point is that the sandbox is not baked into a training script. A team can stand up Orchard Env once, then point three different training frameworks and two evaluation harnesses at the same service.
That decoupling matters more than it sounds. It is the same lesson we drew from the orchestration era of agentic coding: the pieces that survive a re-org are the ones that own a boundary, not the ones that own a codepath. Orchard Env owns the sandbox boundary, and everything else — data collection, RL loop, evaluation, production deployment — becomes a client of that boundary.
The Three Recipes
Each recipe is a full training story with a released model, a dataset, and a set of benchmark numbers Microsoft wants the community to beat.
Orchard-SWE: Software-Engineering Agents
Orchard-SWE trains a software-engineering agent on the Qwen3.5-35B-A3B backbone (about 3 billion active parameters) using a mix of SFT and RL with a resolve-per-rollout (RPR) reward. On SWE-bench Verified it reaches 69.7%, and 73.0% with value-model reranking. That is a new state of the art among open-source methods and closes most of the gap to proprietary systems more than ten times larger. The training corpus is published: 107,185 multi-turn rollouts over 19,287 unique task instances across 2,788 repositories, each with a verified resolve label.
Orchard-GUI: Browser Navigation
Orchard-GUI is a 4B vision-language computer-use agent trained on only 0.4K distilled trajectories and 2.2K open-ended tasks. It reaches 68.4% average success across WebVoyager, Online-Mind2Web, and DeepShop — the strongest open-source computer-use agent to date and competitive with the proprietary systems that ship with browser add-ons. The GUI config includes multimodal per-step rollouts, so the model sees rendered screenshots rather than accessibility trees.
Orchard-Claw: Personal-Assistant Tool Use
Orchard-Claw is the recipe that will interest anyone shipping a mobile assistant or a Hermes-style workspace agent. It trains on only 0.2K synthetic tasks and reaches 59.6% pass@3 on Claw-Eval, then jumps to 73.9% when swapped onto the stronger ZeroClaw harness at inference time. That harness portability is the point: a trained Orchard-Claw agent gains the most of any model compared when its runtime harness improves, which is a direct argument for keeping the agent and the harness as separate release artifacts.
| Recipe | Domain | Backbone | Headline result |
|---|---|---|---|
| Orchard-SWE | Software engineering | Qwen3.5-35B-A3B (~3B active) | 69.7% on SWE-bench Verified, 73.0% with reranking |
| Orchard-GUI | Browser navigation | 4B vision-language | 68.4% avg on WebVoyager, Online-Mind2Web, DeepShop |
| Orchard-Claw | Personal assistant tool use | Compact tool-use model | 59.6% pass@3 on Claw-Eval, 73.9% on ZeroClaw |
Small Backbones, Serious Numbers
Why This Matters for Mobile, Automation, and Agent Orchestration
Orchard's three recipes are not aimed at general chat. They are aimed at three shapes that show up in almost every real product: code, browser, and tool use. A mobile app that has to book a service by walking a supplier's website is running an Orchard-GUI-shaped job. An n8n workflow that opens a pull request against a customer's repository is running an Orchard-SWE-shaped job. A personal assistant that calls three internal APIs to reschedule a meeting is running an Orchard-Claw-shaped job.
The lesson for a Halmob-style stack is not to swap in Orchard's models tomorrow. It is to copy four constraints that the release makes visible:
- Own the environment, not the model. A shared Orchard-Env-style sandbox lets the same agent be trained by one team and evaluated by another without rebuilding the harness twice.
- Ship narrow recipes, not general agents. Orchard's three recipes each target one operational shape. Narrow agents are the ones you can retrain, benchmark, and swap without rewriting the surrounding automation.
- Keep the harness swappable. Orchard-Claw gains almost fifteen points when moved from Claw to ZeroClaw. That is only possible because the trained agent and the harness are released as separate artifacts. Your production stack should look the same.
- Release the trajectories. Microsoft published the SWE, GUI, and Claw rollouts. The reason a product team can iterate on top is that the data is a first-class artifact, not a training-run byproduct.
How Orchard Compares to the Other 2026 Orchestration Layers
The last thirty days have shipped a wave of orchestration layers — Amazon Bedrock AgentCore, Alibaba's Agent Native Cloud AgentTeams, Fenergo Fen-AI for regulated KYC, BridgeApp for the engineering workspace. Orchard sits below all of them in the stack. It is the training and environment layer that the coordinator layers will consume, not another coordinator.
| Layer | What it owns | Who it is for |
|---|---|---|
| Microsoft Orchard | Environment service + trained recipes | Research + product teams training or fine-tuning agents |
| Alibaba AgentTeams | Managed supervisor, policy, identity | Enterprises running many agents on Alibaba Cloud |
| Amazon Bedrock AgentCore | Managed runtime for multi-agent workloads | AWS teams standardizing on a Bedrock-native runtime |
| Fenergo Fen-AI | KYRA agents on a governed record | Regulated banks running KYC/AML with audit-first constraints |
| BridgeApp | Kanban-style engineering workspace with MCP | Software teams running many coding agents in parallel |
For the wider late-July and early-August 2026 context, see our writeups on Alibaba Agent Native Cloud AgentTeams, the Amazon Bedrock AgentCore runtime, and the Fenergo Fen-AI governance stance. The pattern across all four is that the orchestration layer is being commoditized — and Orchard is the training substrate that keeps the agents inside those coordinators from becoming vendor-locked.
What to Do This Week
- 1Stand up Orchard Env against a small sandboxed cluster and pick one real task from your product — a browser flow, a code review step, or an internal tool call — as your first evaluation.
- 2Reproduce the closest recipe. Orchard-SWE if you ship coding agents, Orchard-GUI if you ship browser automation, Orchard-Claw if you ship a personal assistant. Confirm the released model matches the reported number on the released trajectories before doing anything custom.
- 3Swap harnesses. The Orchard-Claw jump from 59.6% to 73.9% between Claw and ZeroClaw is the strongest argument in the paper. Run your own agent under two production harnesses and measure the delta before choosing where to invest.
- 4Publish your own trajectories. The reason Microsoft can iterate at this speed is that the environment and the data are external artifacts. A product team that treats its rollouts the same way is one release away from being able to swap models without rewriting the automation.
Not a Shortcut Past Governance
Where Orchard Fits in the Halmob Stack
At Halmob, most engagements combine a mobile app, an n8n automation layer, and one or more AI agents that keep the two honest. Orchard changes the cost curve on the agent tier. Instead of paying to prompt-engineer a general-purpose model into three different shapes, a team can train one narrow recipe per operational task, keep the environment shared across all of them, and swap the runtime harness later without retraining. That is the shape a real product stack has been waiting for.
For teams designing an agent tier that will outlive one model, our writeups on Hermes workspace mobile orchestration, the executor-advisor pattern, and Gemini managed agents with background execution pair well with this one. Orchard is the layer that makes the agents inside those coordinators cheap to replace.
The Bottom Line
Microsoft Orchard is the 4 August 2026 answer to a question every agent team has been asking since 2024: where does the environment live, and who gets to reuse it? Making the answer open, Kubernetes-native, and shipped with three real recipes is the move that turns agent training from private infrastructure into a public stack. For mobile, automation, and orchestration teams, the leverage is in copying the release's discipline — shared environment, narrow recipes, swappable harnesses, released trajectories — long before adopting the models themselves.
For source material, start with the Microsoft Research announcement, the microsoft/Orchard GitHub repository, the Orchard datasets on Hugging Face, and the smol.ai AINews newsletter for the wider early-August 2026 orchestration-layer context. To wire an Orchard-shaped agent tier into a real mobile and n8n stack — environment, recipe, harness, and audit trail included — Halmob ships that end to end.