microsoft/Orchard-Agentic

Orchard-Agentic is a collection of open-source work on agentic modeling

What it solves

Orchard-Agentic addresses the "train-deploy mismatch" and the lack of portable infrastructure for agentic modeling. It provides a standardized, isolated environment that allows researchers to use the same substrate for trajectory distillation, reinforcement learning (RL) rollouts, and evaluation across different domains (like software engineering and web browsing) without rebuilding the environment for every new study.

How it works

The project is structured into four primary layers:

  • Orchard Env: A Kubernetes-native sandbox service and Python SDK. It spins up isolated containers on demand to handle multi-turn interactions (execution, file I/O, git patches) via a REST API, remaining agnostic to the specific agent harness or trainer used.
  • Orchard Eval: An evaluation layer that runs agent harnesses (e.g., codex, pi) against benchmarks like SWE-bench Verified using the Orchard Env sandboxes.
  • Trainer: An RL training stack (based on a fork of slime) that handles the actual model optimization, utilizing Megatron-LM with SGLang rollouts under Ray.
  • Recipes: Specific research implementations built on the foundation, such as Orchard-SWE (software engineering), Orchard-GUI (browser navigation), and Orchard-Claw (CLI/tool agents).

Who it’s for

It is designed for AI researchers and developers working on agentic modeling, specifically those focused on training LLMs to use tools, navigate interfaces, or solve complex software engineering tasks via RL and SFT.

Highlights

  • Harness-Agnostic: Pre-installs popular agent harnesses (codex, claude, pi, etc.) on the PATH of every sandbox, allowing users to switch harnesses with a simple command change.
  • High Performance: Offers low command-execution latency (0.28s) and supports launching thousands of sandboxes in parallel.
  • Cost Efficient: Significantly cheaper than managed sandbox services, especially when utilizing spot instances.
  • Extensive Datasets: Provides open-source research data including 107K SWE trajectories and 3,000+ multimodal browser-navigation rollouts.

Related

  • Project
    CopilotKit/open-multi-agent-canvas

    An open-source multi-agent chat interface that lets users manage multiple AI agents in one dynamic conversation using Next.js, LangGraph, and CopilotKit.

  • Project
    deerwork-ai/deer-workflow

    A dynamic workflow runtime for building observable, reusable agent graphs using a code-first TypeScript approach to control AI agent execution.

  • Project
    juanjuandog/FinSight-AI

    An open-source A-share equity research workspace that uses hybrid RAG and recoverable workflows to produce evidence-grounded financial reports.

  • Project
    neosigmaai/auto-harness

    A framework for automating the self-improvement of AI agents by using a coding agent to iteratively optimize prompts and tools based on benchmark failures.