Anthropic Engineer Says 90% of Colleagues Are Building Systems That Prompt Themselves; Some in the Replies Aren't Buying It
A tweet about Anthropic's internal agent practices went viral. The video explains how agents remember, correct errors, and get more accurate as they run; some commenters applauded, some saw it as common sense, and others questioned safety.
Anthropic engineer Lamis says in a 30-minute video that you shouldn't spend your effort on prompting Opus 5.5; instead, you should build a system that prompts itself. The original tweet claims that 90% of Anthropic's engineers are already doing this.
The original tweet calls the video more valuable than $1,000–$2,000 agent courses. That claim is catchy, and it's worth unpacking.
What the video is really about is the operating framework for agents: memory, error correction, and getting more accurate after every run. The prompt is just the entry point; the loop happens inside the system.
The concept isn't new. Someone in the replies said outright that this is the most ordinary common sense and they don't know why it's being packaged as a unique idea. Someone else added that memory is the key.
For people who haven't built agent systems, walking through the workflow in 30 minutes is still valuable. Knowing that you need memory isn't hard. What's hard is deciding what to store in memory, how to retrieve it, how to write errors back, how to control costs, and when to stop.
The original tweet doesn't go into these details. The signal it wants to send is that inside a leading lab, agents have already gone from demos to engineering workflows.
Beyond the signal, there's also noise. Some people in the replies joked about Anthropic's system security, saying that even a frontier lab can't protect its own systems; someone compared being an engineer at Anthropic to making chicken at KFC. These are mostly emotional reactions. They don't tell us anything about the technical details, but they expose an open question: the safety, reproducibility, and modifiability of agent systems are still unsettled.
If you want to judge whether this kind of video is worth watching, don't just look at how new it sounds. Look at whether it provides a reproducible architecture, failure cases, and cost data. Without those, even the best slogan is just a slogan.
The video is below. 30 minutes—judge for yourself.
Original tweet link: https://x.com/RamSingh_369/status/2117466916940079498
发布时间: 2026-11-04 03:44