A customer asked for an ALZ design walkthrough on a Monday afternoon. The deck was in their inbox before close of business the same day. Forty-two slides, Fluent design tokens, pre-rendered Mermaid diagrams, accessible contrast, speaker notes, in my own voice. Total operator time at the keyboard: about fifteen minutes.

This session is the stack that made that possible, and how an Azure architect can build the same stack in a week.

The pitch is not "AI is fast." The pitch is that the value of an agentic workflow is in the layers around the model, not the model itself. Three layers, all on stage:

A skill library. Reusable, scoped instructions the agent loads on demand. One skill knows how to build a customer HTML deck with the right design system. Another knows my writing voice in English and Norwegian. A third drives the Sessionize-style CFP format you are looking at right now. I show what a skill file looks like, the gotchas in the loader (the silent drops, the description length limit, the untrusted/ loading convention against the lethal trifecta), and how a small library compounds over a quarter.
A memory vault. A private Obsidian-shaped Git repo the agent reads on every session start. Patterns, decisions, project memory banks, voice profiles, runbooks. Not a wiki anyone updates by hand. The vault writes itself as a side effect of work, with a session-close skill that summarises what was learned and a curate skill that fixes links and merges duplicates. I show the structure, the redaction layer (gitleaks plus a custom content policy plus a git smudge/clean filter so customer names never leave the laptop), and the daily sync.
The workflows that consume both. Live on stage I produce: a forty-plus slide customer deck from a one-paragraph brief, a hands-on lab with a working Terraform module and a step-by-step README, a LinkedIn post in my own voice from a private Teams thread (redacted on the way out), a Sessionize CFP draft that follows the format rules of this exact event, and a one-page architecture review for an ALZ design. Every output is a file I would actually ship.
The talk is not theoretical. Every workflow is the one I use in the field. The skill library and vault that drive the demos are real, currently in use, and the patterns are portable to any architect or engineer who works in Copilot CLI, Claude Code, or any agent that supports skills and a file-system memory layer.

I also show the parts that do not work yet. Skills that hit the description-length ceiling and silently fail to load. Vault context bloat that costs latency and money. The lethal-trifecta risk pattern (untrusted content plus tools plus exfiltration) and the loading convention that mitigates it. What I changed after I got it wrong.

This is a session for IT pros and architects who use Copilot or any coding agent today and want to get past the chat window into a workflow that actually saves hours per week. Not a pitch for any one product. A practical architecture for how to build a personal agentic stack that compounds.