Post

Nothing changed since the 20th century. Waterfall is still broken and AI just makes us hit the wall faster… but still not soon enough to make it a good feedback loop. The instinct is to want faster, smarter agents, or more agents judging each other. 🧵

But I still believe the real move is smaller steps — at least for complex or original features. The real cost is not token usage, and it's not even the review cost. It's what the loop does to you:

- 🛶 The Drift: the slop and unwanted behavior that slips through the review, and whatever your harness is made of, - 🔁 The Context Switching: yes, multi-agent views look cool... but most humans are bad at multi-tasking (I'm among the worst 😅),

- 💰 The Sunk Cost Fallacy: we wait 10 min for the agent, spend half an hour reviewing, and end up... not really convinced. But reverting means throwing all that time away, so we ask the agent to patch on top instead. Now we’re building on something we already don't trust.

The trick is making every step small enough that validating or reverting costs nothing. Around 1 minute to produce. A few minutes to review. And then, every time: continue, or revert and try again.

Cheap reverts close doors to slop. When steps are small, drift has nowhere to hide — the wrong turn shows up in a few dozen lines, not 600. And because you review right after each step instead of babysitting a long run, you stay focused.

Small steps. Tight loops. Cheap reverts. That's the whole game, and none of this is new. It’s just some of the core values of Kent Beck’s eXtreme Programming: feedback, and courage — the courage of reverting.

In the meantime, here's an example where I measured “Common” usage of Spec Kit vs. Charted Coding. Note that on token usage, Spec Kit almost ties, but only if you accept everything and it all goes well. That's the Waterfall bet, and it rarely pays.