most of the time, learning the tools is not the bottleneck.
I use agents for most of my coding now. Claude Code, Cursor — I run these things constantly, and going back to writing everything by hand feels slow in a way that's hard to unsee. But I'm not a power user. I don't memorize every flag, read every changelog, or optimize my setup. And I think that's actually the right call.
There's a version of Sutton's bitter lesson that applies here. His original point was about AI research: hand-engineered approaches lose to general methods that scale with compute. The version for AI tooling is similar — obsessing over the specifics of any one tool is a losing strategy, because the tools change faster than your muscle memory does. Yes, sometimes a feature genuinely changes your workflow. Claude Code's channels did that for me. But those moments are rare, and the time spent hunting for them usually isn't worth it.
The real bottleneck is never the tool. It's knowing what to build, how to break the problem down, and when the agent is confidently wrong. Those skills transfer across every tool and every version. The keybindings don't.
I think people underestimate how much time they spend configuring productivity instead of being productive. There's a specific flavor of procrastination that feels like work — tweaking your editor, trying a new plugin, reading a thread about someone else's setup. It's comfortable because it's adjacent to real work without requiring the hard part, which is thinking clearly about your problem.
The agents will keep getting better. The interfaces will keep changing. The only thing worth investing in is the judgment that sits above all of it.