AI coding agents struggle to fit new feature requests into a product with an intuitive design that feels cohesive. The design and feel of AI-generated components often require significant human steering to be satisfactory.
@levelsio we've been trying to get a version of this feedback -> PRD -> ai agent PR created pipeline running in the last week: https://t.co/J2VqoN9efp the pieces are coming together but the model and orchestration system reliability is not quite there yet. the hardest parts are: 1. helping the AI decide how to fit the new feature request into the product with an intuitive design that feels "cohesive" with the product. the design and feel of the ai generated components never feel quite right without lots of human steering. 2. the ai coding agents themselves are very unreliable right now. we've tried hooking up the feedback + prompts to github copilot workspace but the system has failed for us basically everytime. 3. figuring out where to say "no" or what the most up to date approach is still very reliant on humans. we've tried getting different LLMs to prioritize issues or come up with novel solutions to a specific problems but have anecdotally noticed the ai thinks everything is urgent + the technologies used / approach the ai decides is usually quite out of date. i do think with a lot of polishing across the pipeline and maybe 1-2 model upgrades, this workflow will be very plausible!