Reasoning models help with planning, like R1 + Claude has showed in Aider benchmark: https://aider.chat/2025/01/24/r1-sonnet.html
IMO the key is still code-first agents when we talk about interactions. Although the reasoning models still need to keep getting better at planning to support better requirements what kind of code to create.
Self-developing agent is very possible if it can debug itself out of problems and keeps accumulating solved problems as skills in its skill library.