What the people actually building AI said today. One page — a 2-min read.
Today’s builders focused on what changes when agents do real work: which models power them, how they handle sensitive data, and what happens when they reach a production database. The bigger bet is that easier software creation will expand the need for engineering.
Replit picks OpenAI as Agent’s preferred partner while keeping model choice
Amjad Masad clarified that Replit users can still choose Anthropic and other supported models. OpenAI is now the preferred partner powering Agent because of performance, cost, and the strength of the partnership.
Andy Pavlo says agents need database permissions and safe copies to work on
Pavlo expects agents could eventually drive 10 to 100 times more database queries, though he says it is too early to know whether their query patterns differ much from human workloads. His immediate prescription is familiar database controls: limit what agents can read or delete, and let them test changes on branches of production data. “It's the Wild West for people that don't know the history databases.”
Aaron Levie expects agents to increase the need for engineering
Levie argues that predictions of engineering’s decline assume the world needs a fixed amount of software. As agents make it practical to apply software to more work in banking, life sciences, and manufacturing, he expects engineering to matter more even as manual coding recedes.
Coordinating agents handled a document request for a tax filing
Thibault Sottiaux says a work agent coordinated with his personal agent over email to obtain documents, then the personal agent coordinated with his CPA for a tax filing task. The process took 20 minutes, giving a concrete example of agents working across personal and professional boundaries.
Peter Yang shares a cautious way to audit connected Google apps
With 45 apps and agents connected to his Google account, Yang shared a prompt that reviews every connection and distinguishes basic sign-in from access to Gmail, Drive, or Calendar. It calls for recommendations before any changes and disconnects only apps the user explicitly approves. He also says the Claude Code app has become faster and arguably simpler to use than ChatGPT, while still lacking better browser use, remote control for all threads, and live voice.
Swyx shares the audience numbers behind an AI video channel
Swyx says the channel has served more than 1 million unique viewers a month since May, yet 58% of monthly viewers are new and 68% are unsubscribed. He says recommendations through browsing grew slowly after two years with little traction, while click-through rate remains weak.
Source data comes from the open-source project follow-builders by zarazhangrui, released under the MIT license. Summaries are generated by an LLM from that project's public feeds, and the summarization prompts are adapted from it. Every item above links to its original source.
Summaries generated automatically. Read the original before relying on any claim.
