The Application Layer Isn't Collapsing. It's Just Embarrassing to Look At.

There's a useful infrastructure argument buried in this piece by siddontang — read it for that, not the civilisational framing.

What it actually gets right: AI agent workloads have fundamentally different characteristics than human-driven web traffic. Burst patterns. Machine-speed query frequencies. Mixed transactional and analytical work in the same session, without warning. Most database architectures were designed assuming a human is somewhere in the loop, clicking at human speed. That assumption is quietly breaking. The piece is specific about why, and that specificity is rare in AI infrastructure commentary.

Where it fails is a specific move worth naming.

The diagram. "Human → AI Agent → Database → Result." That arrow between agent and database is doing enormous hidden labour. Every production agentic system at scale has an orchestration layer, a tool-calling layer, guardrails, access control, output validation, error handling, and — for most organisations — mandatory human oversight checkpoints. None of that appears in the diagram. It's not that the author doesn't know it's there. It's that including it would undermine the "application layer is collapsing" thesis, which is also, not coincidentally, the thesis that points directly to TiDB X as the solution.

The complexity didn't disappear. It relocated. That's a real and interesting story. It's just a less dramatic one.

The database infrastructure argument is worth your time. The Andreessen cosplay is not.