The Iteration Loop Got Longer. That Changed Everything.
The thing nobody talks about with AI-assisted development isn't the models. It's the cycle time. The agent's iteration loop got longer, the right way to work changed, and most people are still working as if the loop is two seconds long.
The Progression
Track the iteration loop with AI over the last few years. A clear progression shows up.
- Tab complete: ~2 seconds. Your cycle is essentially synchronous. You type, the model suggests, you accept or reject. Beneficial, but you're still doing normal development. Just typing faster.
- Early Cursor: ~15 seconds. The agent writes a function. You can't do anything useful with 15 seconds, so you sit there, watch it, and the work stays synchronous.
- Cursor with agent mode: ~5 minutes. Long enough that you really want to do something else with the time. Short enough that you can't. The worst place on the curve.
- Cloud agents and slash commands: 20 to 60 minutes. This is the inflection. Now I can do things in between. Now I have something to actually think about in terms of ratios.
METR's measurement work puts numbers on the trend: the length of tasks frontier agents can complete with 50% reliability has been doubling roughly every seven months. That's why the loop kept getting longer, and why it's still getting longer.
What Changes at the Inflection
Once the agent can run for an hour without supervision, the question stops being "how do I keep up with the tool?" and starts being "what's the ratio of synchronous time to asynchronous time, and how do I maximize the parallelism on the async side?"
If a session runs for 55 minutes and I spend 5 minutes per session kicking it off, checking it, redirecting it, theoretically I can have 12 sessions in parallel. Practically nowhere near 12, because most work doesn't decompose that cleanly. But the ceiling moved.
It's like playing chess against 12 different people at once. You're synchronously working the whole time. You're just walking from board to board.
The Prompt-Budget Trade
The way I make this work is by inverting where my time goes. The instinct from short-loop tools is to type a quick prompt, see what the agent does, redirect, refine, iterate. That works when loops are short. It falls apart when loops are 30 minutes.
So I bundle. Spend ten minutes writing a really, really good prompt (aka a ticket), and now the agent runs for 30 minutes without needing me to steer it. This connects to a pattern I've written about before: trying to fix an AI agent mid-chat rarely works, so loading the full intent up front pays back many times over. The moment I kick it off, I spend the next ten minutes writing the next really, really good prompt for a different problem. Kick that off in parallel. Now two long-running things are happening, and I bought the front-loaded planning time by collapsing what used to be three or four steering iterations into one super-iteration.
That only works if the environment supports parallelism: separate work trees, isolated test environments, ports that don't collide. It also only works if your brain can context-switch into a new problem cleanly. Gloria Mark's research at UC Irvine put the average refocus time after an interruption at about 23 minutes. That's the math behind why two-minute loops are so destructive. They interrupt without ever buying enough time to switch usefully. Mine can do the switch in ten-minute increments, not in two-minute increments. So the long-loop world fits my cognition better than the short-loop world did.
What This Means for the Work
The skill that gets rewarded under short loops is fast iteration. The skill that gets rewarded under long loops is precise specification.
If you're still writing prompts the way you did when sessions ran for 30 seconds, short, conversational, "just go figure it out," you're not getting what the new tools actually offer. You're paying for a sports car and driving it like a moped.
The work isn't writing more prompts faster. The work is writing one prompt that's good enough to run for an hour. Then writing the next one while the first is running. That's a different muscle. That's what the longer loop unlocked.
Related Content
