One added benefit is it allows you to throw more tokens to the problem. It’s the most impactful benefit even.

Context & how LLMs work requires this.

From my experience no frontier model produces bug free & error free code with the first pass, no matter how much planning you do beforehand.

With 3 tiers, you spend your token & context budget in full in 3 phases. Plan, implement, review.

If the feature is complex, multiple round of reviews, from scratch.

It works.