Upfront Specification vs Fast Feedback
I'm observing a curious split in how people use AI agents for complex tasks. In one camp, we have those who strive to get all the specifications, context, and prompts just right so the agent can churn away at the task for several hours and return with a perfectly finished task. In the other camp, we have those who strive for an extremely fast feedback loop with small increments and frequent check-ins.
If our experience with software development over the last two decades is any indication, I'll put my bet on the fast-feedback camp. The industry has learned, quite painfully, that a "big upfront design" rarely works. The issues:
No matter how much work you put into the upfront design, there will be deviations during implementation. And a small deviation over a long duration leads to a big deviation.
It assumes that you will learn absolutely nothing during development about what direction you should be going.
The first issue will get better over time as agents grow more capable. The second issue will persist even if they achieve superhuman capability. If you don't take new learning into account and change direction accordingly, faster execution just means running faster in the wrong direction.
Plus, the longer the overall task, the more effort you have to put into crafting the perfect setup, with rapidly diminishing returns. It's much easier to get the AI to do the first little step. And then another little step, and so on. The bounded scope of each step keeps the context uncluttered and lets the AI focus on a small area. That also makes it much easier to review the work!
Of course, none of these ideas are new! That's how effective teams tackle complex tasks, whether or not AI is in the mix. The raw speed of AI just makes the distinction between upfront design and fast feedback iteration much sharper.
