Look, I get it. The “LLMs will only know the boring mainstream stack” worry felt real in 2023. Then I started throwing weird tools at agents and watched them learn the rules in minutes. Simon’s point rings true: the bottleneck isn’t training data anymore, it’s whether you give the agent enough context to become dangerous. That’s a user problem, not a model problem. https://simonwillison.net/2026/Mar/9/not-so-boring/
I’ve seen the same thing building FrameFlow. If I give a coding agent a “go read this repo like it owes you money” prompt, it finds patterns, tests them, and ships. If I don’t, it flails and falls back to defaults. So no, I’m not suddenly going to pick “boring tech” because a model likes it. I’ll pick what fits the job and then force the agent to learn it.
The part that still bugs me is the meta‑bias Simon mentions: not what agents can do, but what they recommend by default. That’s where you get the monoculture—GitHub Actions, Stripe, shadcn, the usual suspects. It’s subtle, but it shapes the stack of anyone who treats the agent like a cofounder. I want the agent to obey, not steer.
P.S. Skills are the quiet power move here. The minute a project ships a good skill, it stops being “boring” and starts being “the obvious choice.”