Definition
Plain language
An open-source framework for training AI agents that separates the sandboxed environment from the training recipe.
As stated in the literature
An agentic modeling framework built on a thin REST-based sandbox service paired with training recipes including credit-assignment SFT and Balanced Adaptive Rollout RL, designed for cross-harness generalization.
Why it matters: Decoupling environments from training recipes makes it possible to train one agent that transfers across many real harnesses.
For example, you can swap mini-swe-agent for OpenHands as the harness without changing the sandbox service Orchard talks to.
Heard on the show
“… The paper is called "Orchard: An Open-Source Agentic Modeling Framework," it went up on arXiv on May fourteenth, twenty-twenty-six, …”Episode 047 — When Agent Benchmarks Lie: The Harness Problem in Open-Source AI