microsoft/Orchard

Orchard: An Open-Source Agentic Modeling Framework

What it solves

Orchard provides a standardized, scalable environment for researching and training AI agents. It addresses the "train-deploy mismatch" where agents are trained in simplified environments that don't translate to real-world deployment harnesses. By providing a stable, Kubernetes-native sandbox service, it allows researchers to develop "recipes" (SFT and RL training methods) that are portable across different domains like software engineering, browser navigation, and personal assistants.

How it works

The framework is split into three primary layers:

  1. Orchard Env: A Kubernetes-native sandbox service that manages the lifecycle of isolated containers. It provides a REST API and Python SDK for command execution, file I/O, and network isolation, allowing agents to interact with a sandbox without needing a specific training stack.
  2. Trainer: An RL training stack (a fork of the slime library) used to optimize agent policies.
  3. Recipes: Specific research implementations (e.g., Orchard-SWE, Orchard-GUI, Orchard-Claw) that use the foundation to train agents for specific tasks using techniques like on-policy distillation and process rewards.

Who it’s for

AI researchers and developers focusing on agentic modeling, reinforcement learning (RL), and the creation of autonomous agents capable of complex multi-turn interactions in software or web environments.

Highlights

  • High Performance: Low command-execution latency (0.28s) and the ability to launch 1,000 sandboxes in parallel.
  • Harness-Agnostic: Pre-installs popular agent harnesses (like codex and claude) in every sandbox, making it easy to switch between them.
  • Multimodal Support: Includes datasets for both software engineering (SWE) and multimodal browser-navigation (GUI) tasks.
  • Crossover Generalization: Agents trained on Orchard Env demonstrate better performance when transferred to harnesses they were not seen during training.
  • Cost Efficient: Significantly cheaper than managed sandbox services, especially when using spot instances.

Related

  • Project
    kubernetes-sigs/agent-sandbox

    A Kubernetes orchestrator for managing isolated, stateful, singleton workloads, designed specifically for AI agent runtimes and reinforcement learning environments.

  • Project
    huggingface/OpenEnv

    An end-to-end framework for creating and deploying isolated, sandboxed execution environments for agentic RL training using a standardized Gymnasium-style API.

  • Project
    alibaba/ROCK

    A scalable sandbox environment management framework designed for agentic reinforcement learning, providing tools for building, deploying, and scheduling isolated runtimes.

  • Project
    octos-org/octos

    Octos is an open‑source, Rust‑based “Agentic OS” that lets you run a full AI‑assistant locally or self‑hosted. It connects to major LLM providers, offers 80+ REST endpoints, multi‑tenant isolation, tool sandboxing, three‑layer memory, and a JSON‑RPC UI protocol. Clients include a web UI (octos‑web), a terminal UI (octoscode), and language bindings (Python, C, Swift, etc.). Install via Homebrew/npm or a script, configure a provider with `octos init`, then start the server (`octos serve`).

  • Project
    opensandbox-group/OpenSandbox

    A general-purpose sandbox platform for AI applications that provides secure, isolated environments for code execution, agent evaluation, and RL training.