November 6, 2025
Agents, chaos, and comment carnage
You Should Write An Agent
Builders say “just make one”; skeptics yell “nobody knows anything” as comments go feral
TLDR: Ptacek urges everyone to try building a simple AI “agent” to learn by doing. Commenters split between skeptics shouting that nobody knows what works and builders sharing multi-agent hacks, while others demand real design patterns and security answers—making this a must-watch debate for anyone curious about useful AI.
Thomas Ptacek’s new Fly.io post “You Should Write An Agent” says AI agents are like learning to ride a bike: you only get it by doing, and yes, the starter code is that simple. The comments? Fireworks. One reader linked a prior HN thread and cheered the hands‑on vibe, but the room split fast.
The loudest camp: skeptics. “Nobody knows anything,” cried one, listing years of hype cycles, copy‑paste projects, and “trust‑me‑bro” RAG (a fancy term for letting a chatbot search your notes). They waved the red flag on hustle culture, calling out grifts and duct‑tape demos. Drama level: high.
On the other side, builders swaggered in with war stories. One dev said context is king and bragged about “agents that talk to other agents,” plus memory and task tracking—basically a to‑do list that argues with itself. Meanwhile, a pragmatist begged for real patterns—think MVC for the web, but for agents—and pleaded, “please don’t say Lang‑anything.”
Then security folks popped in with, “What about the OWASP Top Ten?”—the classic list of web risks—because what’s a party without a vulnerability scare. Verdict: the article says “just try it,” the crowd screams “define it,” and the popcorn keeps popping.
Key Points
- •The article urges developers to build an LLM agent to understand the technology firsthand.
- •It contrasts LLM agents with simpler technologies like the AWS S3 API, highlighting the value of hands-on practice.
- •A minimal Python example demonstrates using OpenAI’s API, a context list, and a simple call function.
- •The code snippet uses client.responses.create with a model labeled “gpt-5.”
- •The piece prioritizes practical experimentation over formal definitions of what an agent is.