The VMware Moment

In 2006, when AWS launched EC2 in beta, a lot of seasoned IT people, me included, clutched our VMware licences and insisted real enterprises would never trust the cloud. In 2008 I sat with a Fortune 500 CTO who said they would never move mission-critical workloads to AWS. What happens when it goes down? I nodded along.

That company now runs 80% of its infrastructure on AWS. I later spent five years at AWS helping enterprises make exactly that move. The irony is not lost on me.

The same fears are back, reworded for coding agents. What if it generates bugs? We are losing our skills. It might leak our code. Some of this is justified. Research from METR found AI tools can increase completion time by 19% for experienced developers on complex tasks, and a rising share of developers say they do not trust the output.

Early AWS was rough too. Instances died without warning. There were no SLAs. I spent whole weekends debugging networking. The companies that stuck with it, learned it and mastered it are the ones ahead now.

The purist and the passenger

The three traps describe what I was doing in 2006. The purist refuses the tools on principle. The judgment stays intact and the scope keeps shrinking, and the argument for the position gets more articulate every year as the ground it stands on gets smaller. I had a very articulate argument for VMware.

The opposite mistake is just as available. The passenger hands the whole job over and keeps the title, and nothing accumulates. Adopting an agent because everyone else has is not the lesson of the cloud. The lesson is that the people who won learned the thing deeply, including where it broke.

Where to start

The developers who thrive will not be the ones who adopt everything blindly or resist everything stubbornly. They will be the ones who know when to let the machine run and when to take the wheel.

That is the delegation ladder, started low. Let an agent write your unit tests first. Then documentation, then the legacy refactoring you have been avoiding. If you run a team, give two or three curious developers a real project rather than a toy one, measure what actually changes, and set clear boundaries on what code the tools can touch and what they cannot.

Each step up is a rung, and each rung should be earned by the answers to two questions: how much of the judgment is written down where the machine can read it, and how expensive is wrong. The cloud taught me that the skeptics were right about the early failures and wrong about the conclusion. Coding agents are on the same curve. The decision that matters is whether you learn them deliberately or get dragged.

Adapted from The Coding Agent Revolution: History Repeating Itself in The Forward View. Read the full issue.