Awesome Foundry Nextgen
Foundry Nextgen. A big bet on Enterprise AI.
A new portal, with all primitives and features required to scale agents under one roof: models, agents, MCP, fine-tuning, governance, safety, security, and observability.
Inspired by earlier ZeroToHero work by Aymen Furter and Hao Zhang, I've built out an extended lab series I wish had existed when I started with Foundry.
In the labs:
- Overviews on the different Foundry pages and describing the control plane.
- "Hub" and spoke foundry account and project deployment patterns.
- Azure policies to control model deployment.
- Multiple agent sub-labs for versioning, code interpreter, hosted agents, memory, MCP, eval, observability, and human-in-the-loop.
- Model fine-tuning.
- Foundry IQ over 3k arXiv papers.
- Guardrails, red-teaming, and safety policies - what they are and how to apply them.
My favourite labs? The two MCP examples. The first is a Project Management Office use case, with 37 tools wired to a custom MCP server on Azure Functions, backed by a synthetic knowledgebase of people, projects, minutes of meetings, risks, and distribution lists. Deliberately an anti-pattern showing what you get when you mechanically wrap a REST API with MCP.
The Contoso Private Banking MCP is designed around intent, not REST, adopting principles from Anthropic's "Writing tools for AI agents" doc, and a counterpoint to the PMO one. 6 tools instead of 37. Designed around what a Swiss private-banking relationship manager actually does in their day, not CRUD operations. Collapsing a workflow, more work done backend-side, fewer roundtrips. The agent gets a tool for the job, not a tool for every step of the job.
