Why most of your AI agent code isn't actually AI
Microsoft just open-sourced the scaffolding that every AI agent team rebuilds from scratch — and its models already rival systems ten times larger.
- 98.4%
- of typical agent code that is infrastructure, not AI logic
- 69.7%
- SWE-bench Verified score for Orchard-SWE
- 68.4%
- average web task success for Orchard-GUI
- ~3B
- parameters, rivaling agent systems ten times larger
On August 3, Microsoft Research released Orchard, an open-source framework for training and running AI agents — software that can take actions on its own, like writing code or browsing the web, rather than just answering questions. Almost every team that has tried to build one hits the same wall: the hardest part isn't the AI itself.
One analysis of production agent code found that around 98 percent is infrastructure — tool permissions, memory management, error handling, and recovery logic — not the actual decision-making AI. On top of that, most teams trained their agents in simplified test environments that looked nothing like the real tools those agents would use in production. When deployed, they underperformed.
Orchard removes both obstacles. It provides the infrastructure layer ready-made and lets agents train directly inside real-world tools — code editors, web browsers, personal-assistant apps — so what they learn in training actually holds up in the real world. Early results are strong: the Orchard-SWE model scored 69.7 percent on SWE-bench Verified, a tough software-engineering benchmark, with roughly 3 billion parameters, rivaling systems ten times larger.
For any team building agents — to automate workflows, handle customer requests, or power a coding assistant — Orchard removes the need to write that infrastructure from scratch. It is open source and on GitHub today. Stop rebuilding the scaffolding; Orchard already did it.