May 14, 2026
Server drama goes durable
LLMs are breaking 20 year old system design
AI agents are exposing old web cracks, and the comments are absolutely feral
TLDR: The article argues that AI assistants don’t fit neatly into the old way websites were built, because they keep state, run longer, and need live back-and-forth. Commenters were split between “this is a real shift” and “please, we solved this years ago,” with plenty of sarcasm in between.
The big claim in this piece is simple: the way websites and apps have been built for 20 years may not work so smoothly when artificial intelligence helpers start acting less like quick search boxes and more like chatty little workers that keep doing things in the background. The author says old-school web systems were designed around a neat rule: store everything in one place, let any server answer any request, and move on. But AI agents want to remember things, keep running for minutes, and talk back and forth live. That, apparently, is where the wheels start wobbling.
But the real fireworks were in the comments, where readers instantly split into camps. One side basically yelled, calm down, we already solved long-running jobs ages ago — comparing this whole “new architecture crisis” to the very old problem of uploading a file and waiting for it to finish. Another camp said AI is not inventing new problems so much as making hidden old problems impossible to ignore. And then there was the pure snark: one commenter roasted the whole framing with “shitty but uncannily useful search,” while another delivered the driest joke in the thread by noting that “durable” appears 13 times. Ouch.
Meanwhile, the more architecture-minded crowd tried to rescue the conversation, saying the author may really be reaching for a system where each AI worker has its own addressable home. In other words: half the room thinks this is a breakthrough, the other half thinks it’s a dramatic rebrand, and everyone agrees the current setup is getting awkward fast.
Key Points
- •The article argues that common web architecture assumes state lives in databases while compute remains stateless.
- •It says LLM and agent systems challenge that model through long-running tasks, stateful process memory, and bidirectional user interaction.
- •Durable execution tools such as Temporal, Inngest, and Restate are described as solving execution reliability but not process routing for interaction.
- •The article states that polling remains a common workaround because standard HTTP and load balancers cannot address a specific running workflow process.
- •It proposes a missing routing primitive based on durable, addressable pub/sub channels rather than connection-based WebSockets.