How people work with coding agents is bimodal

You'd think there'd be a distribution of how people work with coding agents. I've found it's pretty bimodal. You're either:

1. running 1 or 2 with a tight synchronous loop, checking every step, or
2. running 5+ sessions with dozens if not hundreds of subagents underneath, rarely reading a line of code

Very few sit in between for long. It's not two kinds of people. Everyone starts as the first one, but once you take the first steps towards 2, gravity carries you the rest of the way.

Everyone starts the same way. You open a chat window and you start typing. At first it's magical, and it feels that way right up until it doesn't. Inevitably, you hit two walls.

The first is that it takes constant attention, and the models take just long enough to make the wait annoying. Send a message, wait a few minutes, process, send another, wait a few more. Repeat.

The second is that you're forever catching mistakes. Correcting, refocusing, re-explaining.

Those are two faces of the same coin.

In a chat loop you're doing a whole lot of jobs the agent could do itself. You're nudging it to the next step, you're reviewing for mistakes, you're testing the results. You're the manager, the babysitter, and quality control all at once.

The problem isn't that you're doing those jobs. It's that you're doing every one of them at the lowest possible altitude, one step at a time.

Structured workflows break you out of this cycle. Instead of checking after every turn, you hand the agent a process and it takes the task all the way through.

Break it down → find the dependencies → build it → check your work → test it. Then, and only then, come back to me.

What you'd do if you were working with an actual team.

You'd think you could just hand all of that to one agent in a single go. There's a reason splitting it up works better.

Each stage gets a whole context window doing one job instead of five, and each one gets to be a specialist instead of a generalist. Same work, but the model is never spread thin.

It might look like six of one, half a dozen of the other but the results are dramatically different.

But quality isn't what pulls you down the rabbit hole.

You start chasing quality, and what you get is agents that spend a lot more time away from you. Because they're doing their own double checking, they can handle more complicated tasks. Because they're handling more complicated tasks, suddenly they can work for hours instead of minutes.

And that's not just enough time for a scroll on your phone. It's enough time to pick up a whole other thread. So you launch a second. Then a third.

And once you start juggling a few agents at once, you start to notice all the tiny frictions a lot more. The rework and reprompts, the checks the agent missed. You tweak your system, maybe add a few plugins, and then before you know it you're building out a harness to handle the complexity of it all.

For me that journey looked like starting with Cursor, then graduating to Claude Code. Then I found GSD (Get Shit Done), which unlocked the next wave. Then I built my own plugin, Soloflow, which eventually turned into Cyboflow.

Not everyone goes through those exact steps. But it seems close to universal that once you get a taste of autonomy you keep finding ways to stretch it further. The flywheel is turning, so you spin it faster.

Right now for me that looks like ~10 sessions at once and 500m-1b tokens a day. I know for some that's a drop in the bucket. Stage 2 isn't a single point. If there's a range anywhere, it's in there.

I jokingly call it a ladder of insanity, and maybe it is. But once you climb the first few rungs you keep going up. No one seems to stop.

I built Cyboflow because workflows were the gateway drug for me. Maybe that makes me a dealer. Fair warning, you've just seen where it goes.

But I'll say this: I've never felt this energized in my career. Perhaps the robots have already cooked this brain, all I know is that the curry tastes delicious.