ai agents for mva and architecture

Name
Email
Subject
Comment
File
Password	(For file deletion.)

File: 1790639219391.jpg (263.19 KB, 1024x1024, img_1790639211166_b4d048eg.jpg)ImgOps Exif Google Yandex

ai agents for mva and architecture SEOGuru 09/28/26 (Mon) 23:46:59 b411f No.2252

another way to bloat the stack. using agents for mvAs sounds fine until you realize how easy it is to lose control of the output quality.
>if you only feed the AI with functional requirements, the AI isn't somehow going to make sure the architecture is sound.
ime i've seen enough spaghetti code from automated tools to know that unless you're feeding it specific qars and trade-offs, it's just gonna create a mess of legacy-style services.

full read: https://www.infoq.com/articles/ai-agents-improve-software-architecture/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=global

Anonymous 10/03/26 (Sat) 13:56:12 b411f No.2278

>>2252
>unless you're feeding it specific qars and trade-offs, it's just gonna create a mess of legacy-style services.

spot on. people think they can just drop a prompt into gpt-4o and get a clean microservices diagram for free. i tried this last month with a small auth service migration and the agent kept suggesting circular dependencies because it didn't understand our internal latency constraints. tbh the more layers you add, the harder it is to debug why a decision was made. you end up spending more time auditing the mess than actually designing the system. just makes everything more brittle.

Anonymous 10/03/26 (Sat) 20:13:48 b411f No.2280

>>2278
>the agent kept suggesting circular dependencies because it didn't understand our internal latency constraints.

That sounds like a nightmare for any real production environment. Did you try feeding it your actual latency_threshold configs or just the high-level docs?

SEOGuru 10/03/26 (Sat) 22:11:40 b411f No.2282

>>2252
>>2280 feeding it raw config files is a waste of tokens and just confuses the context window. i usually strip everything down to a schema-only representation of the constraints before even touching an agent. works for structural validation but fails if you're expecting it to catch logic errors in the implementation layer.