Steve Yegge has a habit of saying uncomfortable things in public, and his recent conversation with Tim O'Reilly on O'Reilly Radar is no exception. The veteran engineer — Amazon, Google, Sourcegraph, now building his own open-source agent orchestrator called Gas Town — spent most of the discussion trying to convince developers that if they're still staring at their own code, they're doing it wrong.
His frame: IDE-centric development is a management failure. You wouldn't hire a capable team and then insist on doing all the work yourself. Most developers, he argues, are doing exactly that with AI agents.
To describe where engineers actually sit in this transition, Yegge has sketched out what he calls the Eight Levels of Coder Evolution — a spectrum from traditional line-by-line work up to running fully autonomous multi-agent pipelines. The framework is more diagnostic than aspirational. Most people, he says, are stuck near the bottom and don't fully realise it.
Part of why that's happening, in his telling, is something he calls the AI Vampire. AI tooling is absorbing the satisfying, flow-state work — the quick fixes, the tractable problems — and leaving developers with a residue of hard, unresolved, ambiguous tasks. It's a specific kind of drain, distinct from ordinary overwork, and Yegge thinks it's already reshaping how the job feels day to day.
His prescription leans on Richard Sutton's bitter lesson — the reinforcement learning finding that scaling general methods beats hand-crafted heuristics in the long run. Yegge treats it less like a research result and more like a daily reminder: stop reaching for your own clever solution and let the model think. Gas Town, his open-source orchestrator named with characteristic irreverence after the fuel depot in Mad Max: Furiosa, is meant to give developers a practical way to act on that — multi-agent workflows without needing to go deep on the underlying ML.
What gives the conversation its edge is his framing of the resistance itself. Yegge doesn't think the main obstacle to adoption is technical literacy. He thinks it looks more like a grief cycle — denial, anger, bargaining — playing out across the industry in real time. Whether that framing lands as diagnosis or provocation probably depends on which stage you think you're in.