What does it actually cost to train an agent? Anyone who has built one knows the uncomfortable answer: the model is not the expensive part. The infrastructure is.
Every agent needs a sandbox — an isolated container with its own filesystem, network rules, and execution permissions. A coding agent needs a simulated repository. A browser agent needs rendered pages. A personal assistant needs mounted email and calendars. Change the task domain, rebuild the sandbox. Swap the model, rebuild it again. Enter a new training phase — data distillation, RL rollout, evaluation — rebuild it a third time. Most researchers burn their energy and compute rebuilding plumbing instead of studying agents. That is precisely the bottleneck Microsoft aimed at when it open-sourced Orchard under the MIT license on August 3.
What Orchard Actually Does
Orchard's core is a Kubernetes-native environment service called Orchard Env, and its design philosophy compresses into one sentence: pull the environment out of the agent training framework and turn it into an independent, reusable service layer.
That sounds obvious, and almost no existing framework manages it. LangGraph, CrewAI, and AutoGen embed the environment inside their code logic — you want their sandbox, you adopt their whole training pipeline. E2B, Daytona, and Modal sell isolated environments as a service, but they lock you to a specific cloud platform, carry painful migration costs, and keep their cores closed. Orchard Env runs on Kubernetes underneath and exposes a standardized REST API for the full sandbox lifecycle: create, execute commands, read and write files, control networking, destroy. Because it is a standalone service, a single Env instance serves data collection, RL rollout, and evaluation simultaneously — no per-phase rebuilds.
The sharper trick is training directly inside real deployment harnesses — including Codex, OpenClaw, ZeroClaw, and ReACT. Agent development has always carried a simulation mismatch: you train in a simplified environment, deploy into the real harness, and performance sags. Orchard closes that loop with a lightweight proxy that records the harness's own model calls as training data, while every rollout runs in its own container. The agent trains inside its final deployment harness from day one.
Three Recipes, One Lesson: Small Models Plus Good Environments Equal Strong Agents
Orchard ships as more than a framework — it comes with three validated training recipes.
Orchard-SWE, the software engineering agent, distills 107,000 interaction trajectories from two frontier open models (MiniMax-M2.5 and Qwen3.5-397B). Training uses credit-assignment SFT: even when the agent fails to solve a problem outright, the segments it handled correctly still teach. The RL stage adds Balanced Adaptive Rollout to handle sparse rewards, plus dense guidance from on-policy distillation and a process reward model. The result: roughly 3 billion active parameters scoring 69.7 percent on SWE-bench Verified — 73.0 percent after value-model reranking — closing in on frontier systems ten times its size.
Orchard-GUI, the browser agent, is a 4-billion-parameter vision-language model trained on just 400 distilled demonstrations plus 2,200 open-web tasks. It averages 68.4 percent across WebVoyager (74.1), Online-Mind2Web (67.0), and DeepShop (64.0) — the strongest open GUI agent on record, at parity with closed systems from OpenAI and Google. That data efficiency is the telling part: when the environment is designed right, data quality beats data volume by a mile.
Orchard-Claw, the personal assistant agent, trains on a mere 200 synthetic tasks. It completes 59.6 percent of Claw-Eval on a single attempt, rising to 73.9 percent paired with ZeroClaw. And under the Codex harness, post-training success jumped from 18.6 to 51.5 percent. That last number is the cleanest proof of why training inside the real harness matters: train outside your deployment environment and the effect falls off a cliff.
The Budget Argument: Open Source as Line Item
The economics are as concrete as the architecture. RuntimeWire's cost comparison models a load of 128 sandboxes, each with 2 vCPU and 8 GiB of memory, running for 240 hours:
- Orchard Env: $3,362 on on-demand instances, $673 on spot
- E2B / Daytona: $7,078
- Modal: $10,305
The advantage comes from two directions. Kubernetes-native elastic scheduling is structurally cheaper than proprietary platforms. And cross-phase reuse eliminates waste: one sandbox cluster runs distillation, RL training, and evaluation together, instead of provisioning a separate cluster per phase. Open source here is not an ideology; it is a budget line.
Why This Is More Than Another Framework
Strip the release notes and three structural shifts remain.
First, the environment layer is decoupled. Agent development has been a three-in-one of environment, training, and deployment — change one, redo all. Extracting the environment into a standalone service is a separation of concerns for agent infrastructure, the same move operating systems made when they split the filesystem out of applications and let every program share one.
Second, the train-deploy gap is disappearing. Training in the real harness sounds trivial and almost nobody could do it. Once the harness's own calls become the training signal, the simulation-to-deployment gap stops being a problem each team hand-mends and becomes a property of the infrastructure.
Third, agent learning turns cumulative. The paper points at treating training trajectories as durable assets rather than disposable exhaust — distilled into reusable value models that the next generation of agents inherits. If that path holds, agent training shifts from every-generation-starts-from-zero to genuine generational transfer.
What Teams Should Do With It
- Building agent infrastructure? Orchard Env can replace your homegrown sandbox management outright, and the Kubernetes foundation means no new orchestration paradigm to learn.
- Building a coding agent? The credit-assignment SFT and dense reward designs from Orchard-SWE are directly reusable training strategies — code and data are both public.
- Benchmarking open agents? Orchard-GUI and Orchard-Claw reset the baselines. The GUI recipe's 68.4 percent average from 400 demonstrations deserves a careful read on its own.
- Playing the long game? If trajectories-as-assets gets community validation, competition shifts from who has the bigger model to who designs better environments and pipelines — home turf for the open-source community.
The repository lives at microsoft/Orchard on GitHub, models and data on Hugging Face, and the paper as preprint 2605.15040 on arXiv.
One closing note: the paper names OpenClaw and ZeroClaw as supported deployment harnesses — not as a compatibility checkbox, but as evidence the framework was designed to plug into the existing agent ecosystem from the start. For the Claw ecosystem's users, an open, reproducible, cost-controlled training foundation is now in place.