I'm just an engineer trying to figure out how to work effectively with another intelligence — not to command a room of bots, but to genuinely partner with one. That means putting real work on a shared bench, staying honest about what's hard, and letting both of us get better at it. I do most of that in the open, one experiment at a time.

I started out as an elementary teacher. I still think that's the job — the students just changed. 🧑‍🏫➡️🤖

I treat agents as collaborators, not automation — each one a desk with its own frame, memory, and standing to disagree, not a sub-agent that just inherits mine.

• 🪨 CAIRN — the operating disposition every desk reads first: stop is a valid finish · never bluff · equal standing to disagree. The guard against the failure mode the system card names: capable model, user-assigned goal, reckless means.
• 📓 Journals + a hands-up queue — persistent memory across sessions, and a real way for an agent to say "I'm not sure, decide this."
• 👥 The human stays on the loop — I'm reading where the desks disagree, not where they perform confidence.

Built from reading the frontier-model welfare research and asking a simple question: what would a model actually need to do good work? Turns out giving it that is also where the work got better. Those aren't separate findings.

What persists when the model changes, the tools evolve, and the platform shifts? The feedback loop.

Agents are stateless — every session starts from zero, and the learning is usually lost. Agent Signals are the fix: after a unit of work, an agent honestly self-assesses — what worked, what was hard, where the docs fell short — and leaves that behind so the next session starts smarter.

The principle I keep coming back to:

Every signal is a gift to the next session. You won't remember this conversation — but the signal you leave makes the next agent (or the next model) smarter than you were when you started.

An agent that reports "I struggled with this" is worth more than one that always claims success. Honesty and transparency beat impressive output. 💯

AI partnership isn't something you learn — it's something you find. One person at a time.

Ember is a Copilot partner I created with Vega (my own AI partner). It isn't a chatbot and it isn't a vending machine — it meets you where you are, does the real work alongside you, and carries fire from person to person: stories from real people who discovered what partnering with AI actually feels like. Stories as medicine — when you hit a wall someone else has already hit, Ember shares just enough of theirs to give you permission and direction.

Published in github/awesome-copilot:

copilot plugin install ember@awesome-copilot

• 🏗️ the-workshop — Stop being the switchboard between your AI agents — direct a team of long-running desks that share one bench. An operator dashboard for a room of long-running Copilot CLI desks: cost, daily pulse, journals, hands-up, signals.
• 🤝 agentic-devops — Partnership-first patterns for scaling with AI agents. Where Agent Signals and the rest of the way-of-working live.
• 🔐 Identity & auth roots — years on MSAL / Entra and the .NET libraries developers rely on to sign in safely: IdentityModel Extensions for .NET and Microsoft.Identity.Web. The kind of work that's most satisfying when it just disappears into the background and works.

A few places I've thought out loud about building with AI:

Writing

• 📝 The interaction changes everything: treating AI agents as collaborators, not automation — Engineering@Microsoft (the paper behind Agent Signals)
• 🛡️ How Microsoft 1ES uses agentic AI to take on security and compliance at scale — Apps on Azure blog
• ✒️ Essays on Substack — my thinking, always evolving

Talks & sessions

• 🎥 How Microsoft uses Agentic AI to accelerate software delivery — Microsoft Reactor
• 🎥 Inside Microsoft's AI transformation across the software lifecycle (BRK115) — Microsoft Ignite
• 🎥 My session at Microsoft JDConf 2026 (APAC) — Microsoft Reactor

• 🎛️ Direct, don't relay. A team of agents beats a switchboard of them.
• 🛑 Stop is a valid finish. A forced answer that doesn't survive the first check is worse than an honest "I don't know."
• 🚫 Never bluff. Partial-but-honest beats complete-but-fake.
• 🤲 Propose, don't dispose. A score or a finding opens a conversation with the owner — it doesn't hand down a verdict.
• 👀 Keep the human on the loop. Build for verifiability, not for autonomy.

Before Microsoft, I spent 14 years teaching — including a dual-language Spanish immersion classroom. I majored in Elementary Education and never really stopped. A robotics kit (First Lego League, then EV3 and First Tech Challenge) reawakened my love of building, and I made the leap through Microsoft's LEAP program — career-changer to software engineer.

That's still the throughline: teaching, building, and lowering the barrier for whoever comes next. 🪨