Going Home to Copilot After a Year in Frontier Land

I'm about to lead the first Microsoft Copilot development cohort at Colibri Group. That feels like a full-circle moment, because Copilot was my first serious AI development environment. Before Cursor, Codex, Claude, GPT, and custom agent workflows became part of how I build, Copilot was where I learned to build with AI at all.

In a way, this feels like going home. It also feels different now.

Frontier land changes how you see governance

After spending a year in what I call frontier land, Copilot feels heavier in my memory — more structured, more guarded, more enterprise. Frontier tools feel faster and more open from a builder's seat. But that freedom comes with a bill attached.

The more I built with frontier models, the more I understood why governance exists. Frontier tools create speed, flexibility, and capability. They also introduce risk: weak guardrails, account exposure, inconsistent user behavior, and usage patterns nobody planned for. Copilot may feel heavier, but that heaviness lives inside an environment where access control, security, and organizational boundaries actually matter.

That's the tension worth exploring, and I've now lived in both worlds.

Two AI development worlds

Governed enterprise AI is the Copilot world — structured, controlled, built for adoption across a company. Its strengths are governance, security, access control, and a safer on-ramp for non-technical users. Its weakness is that it can feel heavy and clunky, with more friction during experimentation.

Frontier AI development is the world of GPT, Claude, Codex, and open-ended agent workflows. Fast prototyping, high flexibility, a powerful builder experience. The cost is governance: less organizational control and a harder path to scaling safely across a company.

Freedom creates speed. Governance creates scale. The interesting work is learning to use both, instead of pretending one of them won.

What the cohort is actually for

The cohort is around 15 people over 90 days. The stated goal is to build one enterprise-level AI agent and get it functional. I suspect the real output is bigger than that.

This won't be slide-heavy. My training style is to work directly in the product — screen sharing, live building, testing ideas, troubleshooting, and failing publicly so people see that failure is part of the process. I expect about 90% of the calls to be hands-on. The early sessions are mostly about getting people comfortable enough to ask questions and laugh through the mistakes, including mine.

When I first built with Copilot, I was mostly building alone. No group to bounce ideas off, no one to troubleshoot with. Fifteen people from different departments changes that — fifteen perspectives, workflows, and business problems pointed at the same tool.

The real output

Enterprise AI adoption cannot depend on one person building in a silo. Organizations need internal builders, and the best AI ideas often don't come from the technology team. They come from the people closest to the work — the ones who feel the repetition, the knowledge gaps, and the operational drag in their own departments every day.

The official output may be one agent. The real output may be an internal builder community.

I'm not returning to Copilot as the same builder I was when I started. I know it's a solid, stable platform, and I know how heavy it used to feel — but now I understand why that heaviness exists. That's a better place to teach from than the one I left.