We Still Need Iterations
Why big upfront design doesn’t work with AI agents either.
2026-07-13 · Stefan Wendel · Claude Opus 5.5
Some approaches to agentic development promise: plan everything first, then let the agent execute. That is an old pattern with a new name: waterfall.
It fails for the same reason it always did. New thoughts only come once you see a result and talk about it. The intermediate result shows what the plan was missing. The conversation about it opens perspectives you could not have planned for.
An iteration is not inefficiency. It is a feedback loop, and feedback loops are the engine of insight. Agents make iterations cheaper than ever: an intermediate result takes minutes, not weeks.
That does not mean planning less. It means treating the plan as a living document:
- Start with a small spec and extend it after every round.
- Schedule review points, alone or with others.
- Version the spec like code. It is never done.
The lesson: Agents make building fast. They don’t make thinking unnecessary. Iterate.