The shield, not the whip

A colleague and I were talking over lunch today about how coding agents have changed the shape of our work. Not the volume — the shape.

It used to be roughly 60% writing code and 40% making decisions. Which abstraction to use, how to name things, where to draw module boundaries, what to build next. The coding filled the hours but the decisions were the real work.

Now the agent codes faster and often better than I do. It doesn’t get tired, doesn’t lose focus, doesn’t need to look up API signatures. So the coding part has compressed to near-zero, and what’s left is 100% decision-making. That sounds like a promotion. It’s not.

When you’re coding, there are long stretches where the work is mechanical. You know what needs to happen, you just need to type it out. That’s rest for your brain. The agent took away the rest periods and left you with a continuous stream of judgment calls. Accept this diff. Reject that approach. Rewrite this prompt. Scope that feature. Every single one costs mental energy, and there’s no downhill stretch where you can coast.

Eight hours of coding with decisions sprinkled in is tiring. Eight hours of pure decision-making is exhausting.

And here’s the uncomfortable part. If an agent makes you 10x more productive, you can work 8 hours and produce the output of 80. Your company wins. You burn out in three months.

Or you can work fewer hours and produce the same output as before — but your competitor picks the 80-hour option and eats your lunch. Neither choice is great, and most companies will drift toward the first one without explicitly choosing it, because the pressure is invisible. It just looks like “keeping up.”

I’ve been feeling this with my own tools too. I built Grove, a little CLI for managing worktrees across repos. I use it daily, it makes me faster. But the other day I realized — it’s a whip. It removes friction, automates steps, lets me context-switch at speed. More decisions per hour. More cognitive energy burned. Most developer tools are whips. Faster builds, faster deploys, faster code generation, hot-reload loops. They all compress the mechanical parts and leave you face-to-face with the decisions, sooner and more often.

What I actually want is a shield. Something that absorbs cognitive load instead of generating it. Something that reduces the number of decisions I need to make, or gives me enough confidence in a decision that I stop spending energy second-guessing it.

Shields aren’t a new idea. We’ve had them forever, we just didn’t call them that. IntelliJ is a shield — it catches your mistakes and suggests the right method before you even ask. You could write Java in Notepad, but IntelliJ absorbs a thousand small decisions for you. Trello is a shield — instead of holding every task in your head and worrying about what you’re forgetting, you offload it to a board. The cognitive weight just lifts.

So the shield for the agent era — I’m not sure anyone knows what it looks like yet. But I know the shape of the thing we need, even if the thing itself doesn’t exist yet.

The conversation over lunch ended without a neat conclusion, which felt honest. We’re in the middle of this shift and nobody has it figured out. The bottleneck isn’t typing speed anymore. It hasn’t been for a while. The bottleneck is the number of good decisions you can make before your brain checks out for the day.

The agents are going to keep getting faster. I want to build the shield. I just don’t know what to build yet.