Architecture in the era of AI: quotes and takeaways from Ilya Klimov's Fwdays talk

In his talk 'Architecture? Never heard of it' at JavaScript fwdays'25, Ilya Klimov raised a question that every senior software engineer and tech lead is asking today: what happens to software architecture when AI models generate 80% of our code?

You can watch the full speech on YouTube: https://www.youtube.com/watch?v=EB6wd7SiZsc&t=2329s - particularly around timestamp 38:49 (2329s), where Ilya summarizes his reflections on AI therapy and vibe-coding in production.

Here is a key quote from Ilya Klimov at timestamp 2329s: 'Architecture did not die with AI — in fact, it became the engineer's primary lever. Earlier, we built complex layers of abstractions to make developers' lives easier. Now, we must build transparent, verifiable systems that AI agents can navigate and humans can trust.'

He emphasizes that vibe-coding without architectural boundaries inevitably leads to unmaintainable complexity. Generating code is cheap; understanding, verifying, and maintaining context is where real engineering happens.

From my perspective as a Senior Full-Stack & AI Engineer, Ilya's message hits the nail on the head. LLMs excel at generating localized logic, functions, and boilerplate. But without explicit system boundaries, typed contracts (Zod, TypeScript schemas), and deterministic guardrails, AI-generated code devolves into spaghetti fast.

The shift isn't away from architecture — it is toward readable architecture. Clear contracts, modular components, atomic tools, and strict validation are not just good practices for human teammates; they are prerequisites for AI agents to reason accurately.

If you are building products with AI integrations, RAG pipelines, or agentic automation today, remember: the model is a component, not the system. Clear architecture is what turns vibe-coded experiments into production-grade software.