Designing AI-Intensive Applications - swyx

A useful AI product can be a carefully arranged workflow. In this June 2025 talk, swyx describes the AI News system as scripts that gather source material, plan the processing, summarize it in stages, format the result and evaluate it.

The design breaks a large reading task into smaller pieces, processes them and combines the results into something a person can use. The talk proposes this as a pattern for applications that perform substantial AI work behind a small amount of user input. The relevant design question, in this account, is whether those stages produce valuable output; counting model calls or debating whether the system qualifies as an agent does not answer it.

“We used to be low status. People just derive GPT rappers and look at all the GPT rappers. Now all of you are rich.”

swyx

“what is the standard model in AI engineering right we have standard models in the rest of engineering”

swyx

“he reads AI news he loves it but it is not an agent. And I was like what do you mean it's not an agent?”

swyx

“we want to deliver value instead of arguable terminology”

swyx

“when you start to make real money from your customers is when you start to do evals and you start to add in security orchestration and do real work”

swyx