My thoughts thus far on agentic development, and its implication on the future of software engineering
Sometime in January of 2026, the raw capability of LLMs in relation to software engineering crossed a critical threshold. Until that point, I had not primarily been using LLMs to generate code. Instead, I would write the code “by hand”, but when faced with a specific syntax question or (gasp) a need to craft a SQL statement, I would shoot off a question to my favorite LLM. I had tried using LLMs for end-to-end code generation, but found the tradeoff not worth it. While they generally produced code faster than I did, they produced code of a low quality: poorly organized for sure – monstrous methods and plenty of “run-on sentences”, but also more often than not with obvious and non-obvious bugs. Thus, completing a ticket often took about the same time with and without an agent writing the code. ...