I’ve been using LLMs for writing code since late 2022, starting with the Copilot VS Code extension. I’ve seen improvements to my own output early on but always steered clear of the hypetrain and never bought into outrageous claims of 100x productivity improvements, simply because they didn’t reflect my own experience. Those claims for early models were obvious hyperbole in hindsight, and Github Copilot ca 2022 was in fact little more than smart code completion. But things have changed qualitatively over the last couple of months as agents finally started to work in 2025. The current generation of models — and importantly, advances in context engineering and the tooling built around them — are starting to fundamentally change…
I’ve been using LLMs for writing code since late 2022, starting with the Copilot VS Code extension. I’ve seen improvements to my own output early on but always steered clear of the hypetrain and never bought into outrageous claims of 100x productivity improvements, simply because they didn’t reflect my own experience. Those claims for early models were obvious hyperbole in hindsight, and Github Copilot ca 2022 was in fact little more than smart code completion. But things have changed qualitatively over the last couple of months as agents finally started to work in 2025. The current generation of models — and importantly, advances in context engineering and the tooling built around them — are starting to fundamentally change the nature of writing software, and therefore what it means to be an engineer.
Do I still wish we had 2x as many engineers on the team? Absolutely. But the engineers I want now are different from the ones I would’ve hired for a couple of years ago. For a long time, you could make a good living doing pure implementation — knowing the ins-and-outs of your chosen language, every trick in the SQL book, etc and using that to write optimized implementations based on unambiguous specs. While that continues to be relevant, I suspect the amount of value you’re going to be able to capture from a pure implementation focus are going drop fairly quickly from here. Agents are a multiplier on your software and system design skills, less so on your implementation skills. What matters now is writing clear specs and requirements (and writing clearly and concisely in general), having strong intuition for the right abstractions, and a deep understanding of what drives codebase complexity and how to keep it in check. Demand for those skills has always been high but it’s going to explode in the coming years.
It’s a classic case of commoditizing the complement. Implementation — writing code — is an economic complement to specification — writing down what the code should do. They’re strong (but not perfect) complements since each on their own isn’t worth nearly as much without the other. You can write code that does things but without a clear spec, it’s unlikely to deliver good value. Likewise, you can write the perfect spec but if you can’t implement it — within time and resource constraints — it remains an academic exercise. Coding agents are commoditizing implementation, leading to an increase in demand for its complement, writing specs.
Making software is insanely costly today and that cost is almost entirely driven by personnel. Historically, quality software has always been supply- not demand-constrained. There’s no obvious limit to how much good software we need in the world. Commoditizing one of its key supplies is a really exciting development.