Skip to content

What's new

Reference · 5 min · no code

What changed on this site, newest first.

Changes to the AI Knowledge Hub, newest first.

September 2026 — enterprise guidance

Three pages of best practices for teams building agents and MCP servers inside an organisation. They are written as numbered rules (MUST, SHOULD, AVOID) that a design review or pull request can cite, and each rule links to the incident or standard behind it. They are framework-neutral, with Azure named where a control has an obvious home.

  • Identity and authorization covers the four ways an agent runs (for a signed-in user, in the background as itself, in the background for a user who has left, and as its own user account), and why no token should travel further than the service it was issued for.
  • Building agents covers design-time risk checks, limits enforced in code, containment for agents that run code, human approval that people still read, and release gates.
  • Building MCP servers covers tool design, what to do when a catalogue grows past what a model can choose from reliably, tool integrity, and supply chain.

September 2026 — one ramp, two paths

The site had grown two halves that did not meet: a set of concept pages written in March, and a hands-on agent path added in August. Nine top-level tabs held them side by side, and nothing routed a reader from one to the other.

It is now organised as four layers, and most topics appear at more than one of them on purpose.

  • Choose your path is the new entry point. Two routes: a ten-page reading path with no code, and the build path with ten labs.
  • Six tabs instead of nine, ordered as a ramp: Start here, Understand, Build, Go deeper, Reference.
  • The reading path puts retrieval before agents, because giving a model your data is a smaller step than letting it act, and safety after agents, because the failure that matters arrives through a tool result.
  • Every page is tagged by layer and topic, and Browse by tag shows a single topic across every depth.
  • A new build module: structured output, between tool calling and the agent loop. It measures prompting, JSON mode and schema-constrained decoding on the same task, and shows the failure a schema cannot fix: asked for a required field the message never mentioned, the model invented one on every run, on two different providers, and every invention was schema-valid.
  • Two more build modules, both measured rather than asserted: memory, where one inferred fact written to a store was reported as established truth in every later conversation, and multi-agent, where splitting a task cost 2.6x the tokens and produced a worse answer because a specialist judged its own document irrelevant.
  • The labs run on Azure OpenAI, verified end to end. Setup and the labs readme now carry a provider tab for it.
  • The six C# samples are on Microsoft Agent Framework 1.x and now run, against Azure OpenAI with the same three environment variables the Python labs use. They previously targeted Foundry server-side agents, which needed a project and an az login, so they could be compiled but seldom run.
  • No URL changed. Every page that was reachable before is reachable at the same address.

August 2026 — the build path

Ten modules, each ending in code that runs against a model on your own machine. No account, no API key, no cost.

  • The build path and Setup, covering four hardware tiers.
  • Nine modules: tool calling, the agent loop, the harness, context engineering, retrieval, evaluation, observability, safety and production.
  • Ten labs in the repository, standard library only, each one demonstrating the specific failure its module is about.
  • Resources — a ranked reading list with a reason to trust each item, and an explicit list of what to avoid and why.
  • Corrections. The Model Context Protocol page was rewritten against the 2026-07-28 revision. The Copilot CLI and Extensions page was removed: both products it documented were retired in late 2025. The frameworks page was rebuilt around Microsoft Agent Framework, which superseded Semantic Kernel and AutoGen in April 2026.

March 2026 — launch

The site went live with the concept, pattern and reference sections: AI 101, eight concept pages, RAG and knowledge systems in six parts, architecture patterns, developer tools, the glossary and official sources.

The original "Learning Paths" page offered four routes by role. It was replaced in August by the single build path, and in September by the two paths on Choose your path.

Corrections and additions are welcome: open an issue or a pull request on GitHub.