Has anyone else noticed that AI get dumber as the context window gets larger?

[Mharbulous]

I have been trying to use the BMAD-METHOD to rebuild an app that I previously built with AI coding without any real method. With my original app, as the complexity of the app increased the AI seemed to get worse and worse at coding until eventually progress became so slow that I basically stopped. Unfortunately, using the the BMAD-METHOD to rebuild the app from scratch, the same thing happened

With either approach, the BMAD-METHOD or the dialogue approach, eventually the app seems to reach a point of complexity where the AI starts getting worse at coding, and progress becomes slower and more expensive. The fundamental issue seems to be that the AI gets dumber as the context gets larger. It doesn't seem to matter whether the complexity is well documented, (as it is with the BMAD-METHOD) or barely documented (as with my original attempt). Am I doing the method wrong?

Has anyone else had the same experience?

[OverlordBaconPants]

I have found the BMAD method keeps context windows smaller by focusing efforts only on one atomic, discrete Story at a time. When the Acceptance Criteria have been met (including updating corresponding documentation), then start a new conversation.

I have also found that I need to close out all open tabs in my IDE before starting any conversation, lest the AI agent focus unnecessarily on those tabs.

[kuldeeprajput-at]

Exactly i actually agree with that because if your context is bigger then you are surely doing something wrong
first we need to learn how to provide context in chunks, and then we can start diff diff conversation tab with that we can surely achieve what we want.

[bmadcode]

Yeah - I think that will happen regardless of method. The reason I think BMAD method can help is it shards and creates smaller reference docs and almost self contained stories so the agent should know directly what to do. But as mentioned - the method really stresses, create a story -> start a new chat with dev and implement -> start a new chat with SM and create story -> start a new chat with dev and implement, etc...

BUT - its also helps regardless of method to follow some basic red-green-refactor, or refactor between stories, strive to keep files small etc...

Anyways hope this helps a bit - will try to share some more specifics when I get a chance to create another video. Would love to know if any of these suggestions help! Good luck @Mharbulous :)

[SWALK10]

That is a known problem with agentic AI processes. Can only have a limited rounds of interactions and context, no matter how much instructions, docs, mds and rules they have, and after about 3-4 rounds they start to decay, first slowly, then all at once they self-destruct with your code.