I’ve spent about a billion tokens in the last two months, all through a combination of terminals, desktop clients, and IDE extensions. I’m borrowing the format from Ben Tossell, who counted three billion in four months in his post on coding with agents.
I am now watching agents write code I couldn’t write myself. I have written PRDs and implemented complete flows for a cybersecurity platform, built AI evaluation engines for compliance and product validation, and run multi-agent systems that critique design output from other agents or from humans.
Some people call this “vibe coding.” I think that undersells it. This isn’t about vibes. It’s about having developed a clear judgment from the criteria we build empirically as humans, and about learning to orchestrate systems that produce quality at speed. The skill isn’t writing code. It’s directing intent, managing context, and knowing when to step in. That’s what’s becoming more aligned with what people call “taste.”
Here’s what that looked like the week before I wrote this. I started NIS2.chat on New Year’s Eve. Before the agent wrote a component, I gave it the context it needed: the product, the stack, and the workflow it had to follow. Every task started as a written plan, tests came before the first component, and it built the layout, the chat window, the input, the state store, and a mock API in one sitting. A few sessions over the new year and I had a working chat app, including a restyling pass of my own when the pages drifted from the spec.
As a product designer, I’ve always been more interested in what gets built and why it should be built than in how it gets built. But the how has changed forever. And I’ve changed with it.
Also in January: What I Actually Built with Agents · A New Kind of Technical Career · Learn the Loop, Not the Tool
Daniel Arevalo is Product Experience Lead at Omny Security. He builds AI-native tools and writes about design, product leadership, and what it means to direct systems that execute at a speed humans never had before.