What they did
Vankayalapati, a Dynamics 365 Finance and Operations developer, wrote a team of eight agents (an architect plus finance, supply chain, security, reporting, developer, reviewer and QA specialists) and runs them in a Code session of the Claude desktop app on a D365 repo. They sit on d365fo-mcp-server, an open-source MCP server by Filip Manas that indexes the real metadata on the dev machine and lets the model search objects, validate code, build and run the best-practice check. The user only talks to the architect, which decides which specialists a ticket needs. Each agent's tool list is its permission list, so only the developer can write X++. PowerShell hooks outside the model record a sign-off tied to the exact design version and block any metadata write or build without it. Every step leaves a file in the repo: assessments, functional and technical designs, build log, review, QA result and a UAT script.
How it went
On a short test ticket asking that purchase orders for vendors on hold not save, the architect found the standard code already refused the save and only warned, so it proposed a smaller fix: a Chain of Command extension on the account check, one label and four SysTests. After the author typed go, the developer agent wrote and built them, and all four tests passed in Visual Studio on the first run. The agents are re-run against 23 sample tickets scored on eight checks. Database sync, deploy and test runs still happen by hand in Visual Studio.
Worth knowing
This is the author's launch post for his own free, MIT-licensed project, and it was only verified on UDE environments in the Claude desktop app; cloud-hosted VMs and GitHub Copilot are untested. A full ticket burns far more tokens than a chat answer. The approval hooks match commands as text, so they stop mistakes, not a user set on bypassing them.