Agentic Engineering & The Great Extrapolation
Last year, there was a marked change in the programming capability of the latest Large Language Models (LLMs). AI-generated code became professionally usable, and in combination with the release of tools such as Claude Code, Agentic Engineering was set to take off in the work place.
How has it played out?
Ideation, rate of iteration, initial design, debugging and Mean Time To Recovery (MTTR) have all improved. Since Thariq Shihipar's tweet on the effectiveness of HTML outputs from these tools, I've been approaching problems differently too, reducing decisions to multiple-choice, then often refining these choices toward an optimal solution.
One of the wonderful things about LLMs is that they do not get bored. Where I fatigue two hours into fixing a sticky bug, AI just keeps going. It's immensely satisfying, too, turning round a task in an afternoon that previously would have taken me a week.
However, it's messy. Engineers are handing out 10,000 line PRs like they're going out of business, the most successful SaaS providers are grappling with demand, and Junior Engineers are not being given the opportunities to learn. One METR study highlights a bizarre phenomenon where experienced engineers believed they had completed tasks faster using AI tools, when in fact doing so had slowed them down. Clearly, it's sub-optimal to reach for these tools as your default.
Do we even need engineers?
Like many others, I wondered whether I'd be in a job following …