How to maximize the power of planning when building AI agents

I will share what I’ve learned through the process of building my startup - codapt.ai. The goal is to allow vibe coders build solid production-grade apps. Codapt has went through ~16 iterations as LLM models kept involving since 2023. During the process, I tested different way of context population, planning, subtasking, etc. to squeeze everything out of LLM models to strike a good balance between code quality and latency. From our users’s feedback - Codapt generates better code than major vibe coding tools on the market. I can share my first-hand learnings on agent pipelines, best architecture for planning and what I believe the trend would be for AI agents.