/
Published on 11.03.2026
TLDR: Steve Yegge, veteran engineer from Amazon, Google, and now Anthropic, lays out an eight-level framework for AI adoption among engineers, argues that the IDE is evolving into a conversation and monitoring interface for AI agents, and warns that AI-augmented work creates a "Dracula Effect" that drains engineers faster than traditional coding.
There is a conversation happening right now that every software engineer needs to pay attention to, and Steve Yegge is one of the most interesting voices driving it. On the Pragmatic Engineer podcast, Steve sat down with Gergely Orosz to talk about how AI is fundamentally reshaping the craft of software development. Not in some abstract, hand-wavy, futurist way, but in the practical, "this is happening to your team right now" kind of way.
Steve has this framework of eight levels of AI adoption. Level zero is the engineer who refuses to touch AI tools at all. Level one is asking an IDE for suggestions and carefully reviewing the output. And at the top end, you have engineers orchestrating multiple AI agents running in parallel. The uncomfortable truth Steve points out is that most engineers are still clustered at levels one and two, and he suspects those who stay there will eventually be left behind. That is a provocative claim, but when you look at what teams at Anthropic are doing, shipping prototypes to production in ten days through what Steve calls "slot machine programming" where you build twenty implementations and pick the winner, the gap between those approaches is staggering.
One of the most fascinating observations is what Steve calls the evolution of the IDE. He sees tools like Claude Cowork as signaling a fundamental shift where the IDE stops being a code editor and becomes a conversation and monitoring interface. You are not typing code anymore. You are directing agents, reviewing their output, and managing workflows. Gergely pushes back on this a bit, noting that conversational tools have not fully arrived yet, but the trend line from tools like Claude Code becoming wildly popular among developers is hard to ignore.
Steve also drops what I think is one of the most underappreciated insights in this whole conversation: monolithic codebases are a major blocker to AI adoption in enterprises. AI agents have an effective ceiling somewhere between half a million and a few million lines of code. If your codebase is a monolith that will not fit in a context window, AI agents simply will not work well with it. For architects and team leads, this is a concrete, actionable signal. If you have been looking for the business case to break up that monolith, AI adoption might be the strongest argument you have ever had.
The "Dracula Effect" concept is genuinely brilliant and something every engineering manager needs to internalize. When AI automates the easy tasks, engineers are left doing nothing but high-intensity thinking all day. Steve estimates you might only get three productive hours at maximum speed, but during those three hours you could produce a hundred times more output than before. This has profound implications for how we structure engineering workdays, how we estimate capacity, and how we prevent burnout. The old model of eight productive hours at varying intensity levels is simply not how AI-augmented work functions.