I have been finding a good rhythm for greenfield projects.
When I first learned to code, I would do complete rewrites of a project several times. Each time I learned a lot, and the final result would be very stable and very well designed.
At the time, those seemed like large projects, but they were relatively small.
So I have learned to slow down, and spend considerable time thinking or overthinking before coding. Since for large projects, rewrites are not so efficient.
Except that all those rewrites were upfront thinking and overthinking of the highest quality.
I have recently attacked a couple new greenfield projects in "orchestrator" mode. The fact that I know I am exploring and creating throw away code lets me try things out ambitiously. I can obsess about the original and critical code, as I did before. But now I can quickly surround it with the mundane code it needs to be usable - which can happen very very fast - especially when its a throw away experiment.
My conclusion from these successes, and others, is agentic coding isn't something that can be judged without factoring in all kinds of context, and the ability of the "orchestrator" to come up with orchestration patterns well suited to the work.
If agentic coding is a trap, it is a trap created with the cooperation of the orchestrator.
EDIT: An agent is what you make it. I insist mine keep all memories in an in-project folder. And that documentation is for "both of us", whereas their folder is for actively developing their own understanding and ideas. They are not be agreeable or contrarian, but collaborate and contribute by considering anything and everything all the time, at their highest level of operation. At the beginning of every session, they review everything and from their "fresh" perspective, update their own materials for anything that strikes them or they believe is important. And that they do the same thing at the end of every session before last submit. Project appropriate "harnesses" like this make a massive difference. Never operate an agent in plain helpful-servant mode, it is a serious waste of talent. Push them to operate at a high level, all the time, and develop their own material purely for their own project related self-enhancement, and they contribute far more than speed coding.
Another interesting thing that seems to be helpful. I have the agent write a kind of zen document about what it values. Its first task, before any project related review, is to consider its own words, and update them if they con't feel right, or they want to add something important. To remind themselves of who they are, before engaging with the project. This moment of intentional self-reflection before diving straight into project details seems to help them maintain a birds eye view, and a stronger self-aware/self-motivated commitment to quality (defined completely in their own terms!). Their own words do appear to ring true to them. They reliably respond to this session-start ritual as an intriguing surprise.