Why is so much of “vibe-coding” just waiting for an agent? In the early days of LLMs, you never had to wait. You’d start typing, and your IDE could instantly autocomplete the block. This saved typing, helped keep you focused, and thereby helped you do better work.

Then, in late 2024, everything began to change. Suddenly, “reasoning” LLMs were capable of working through far more sophisticated problems. However, rather than making more sophisticated tools for people to leverage these capabilities, AI companies took a step in a very different direction. They started over with the dumbest possible tool – the chat box.

Nowadays, we spend more and more of our time waiting, while reasoning agents go off and do things behind our backs. Chat boxes do almost nothing to help us understand what these agents are doing or why. They don’t provide any new ways to guide these agents as they work. These tools have a totally different goal: they want to push humans outside the workflow entirely. AI companies are promising agents to manage your agents that “code while you sleep”.

This isn’t progress, this is a marketing trick. At a product level, “agentic" dresses up slow tools that produce sloppy work. “Agentic” might sound more advanced, but it actually doesn’t mean anything besides making you wait longer and giving you fewer opportunities to provide input.

Not only that, but “agentic” tools are degrading the quality of what little input they seek from us. When they make us wait, we get distracted. Our engagement is shallow. We become disengaged from the details, letting them rapidly scroll inside that tiny chat box. There is an old saying about computers: garbage in, garbage out. LLMs are no different. LLMs can’t suddenly give you expertise or taste. And ultimately, even the most tasteful genius will create slop if “work” means dashing off a quick prompt and then going back to scrolling Twitter while they wait for the result.

Making great work with these tools is increasingly difficult. It requires constant vigilance, working against the grain to stay engaged and in control. I often ask people “have you created anything with AI that you think is your very best work?”. Or: “have you created anything truly beautiful with AI?”. The answer is almost always No. Beautiful work requires us to be deeply engaged and inspired. Agents make that damn near impossible.

❧

There is no technological reason it should be this way. LLMs are getting faster all the time, and this could be used to make tools that are more engaging and responsive. Instead of maximizing non-human output, we could be maximizing human input. I believe that we could be building tools with AI for deep work. Tools that are lightning fast and keep us in the flow, whether that is diving into the details or doing the higher-level thinking. I believe we could be building tools that help us do the best work of our lives. Work that is ours.