Swyx, affiliated with SMOL AI, Cognition and Latent Space
Swyx shares audience figures for his channel: it has served more than one million unique viewers in almost every month since May, following AIE EU. He says 58% of each month’s viewers are still new and 68% are not subscribed. Viewing is concentrated between 7 a.m. and 11 a.m. Pacific Time, with little activity after 5 p.m. He recalls roughly two years of limited attention before browse recommendations gradually appeared, although click-through rates remain poor. He also wants to collaborate with the other creators his audience already likes, specifically mentioning Jeff Delaney and ThePrimeagen.
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux describes coordination between his personal and work agents. At the request of his personal dot, “Dorito,” his work dot, “Burger,” coordinated over email to obtain documents. The personal dot then coordinated with his CPA on some tax filing matters. He says the process took 20 minutes. The post does not specify the tax matters or the final filing outcome.
AI tutorial and interview creator Peter Yang
Peter Yang found 45 apps and agents connected to his Google account and shares a prompt for auditing them. It first checks every connection category and asks which apps the user still uses, rather than guessing. It then opens each app’s details and distinguishes basic Google sign-in from permission to access Gmail, Drive, Calendar or other data. He asks for separate recommendations to keep, disconnect or ask about each connection, with reasons, possible breakage and flags for broad permissions, unfamiliar providers and duplicates. Recommendations must be shown before any changes; only individually approved connections may be disconnected, followed by verification, without deleting app accounts or cancelling subscriptions. In a separate experience report, he finds the Claude Code app faster and simpler to use, while still wanting better browser and computer use, remote control for all threads and live voice.
Replit CEO Amjad Masad
Amjad Masad clarifies that Replit remains model agnostic. Users can choose Anthropic models or other models the platform supports. His remarks at TC Disrupt concerned OpenAI becoming the preferred partner powering Agent. He cites performance, cost and the strength of the partnership as the reasons. His stated aim is to preserve users’ model choice while optimizing the performance and price of the default experience.
Box CEO Aaron Levie
Aaron Levie argues that predictions of software engineering’s demise assumed the world needed a fixed amount of software, so automation would reduce demand for labor. He expects the opposite: agents that use tools, operate computers, write software on the fly and process data make previously impractical digitization possible. In his view, this could expand engineering’s role in the economy. His examples include new forms of bank trading, research and laboratory work in life sciences, and manufacturing supply-chain automation. He expects less manual coding but more systems and processes to become code, preserving the importance of people who understand them. This is his view of how work will evolve; the post supplies no employment forecast.
The MAD Podcast: Andy Pavlo on databases in the agent era
**The first database changes agents bring concern traffic volume and combinations of reads and writes; permissions, isolation and business semantics still require careful work.** Database researcher and Carnegie Mellon professor Andy Pavlo, who joined ClickHouse to establish a research lab, discusses the shift from agents using databases to writing and tuning systems. First, database branches provide copy-on-write snapshots of production data. They let coding agents test application and schema changes in isolated copies before deciding how to deploy them. Pavlo stresses that controls over reads, writes and destructive operations already exist, and agents should not receive a single set of unrestricted company credentials. As he puts it, a system needs to understand “who’s asking for it, what are they asking to do, and should they be allowed to do it?” MCP offers a common interface; it does not replace those checks. Second, agents can run around the clock, with one person potentially delegating to many of them. He estimates query volume could increase by 10–100 times, but emphasizes that this scale is not yet widespread. He also does not know whether agents will generate fundamentally different complex queries. Traffic volume, read/write combinations and operational telemetry need separate observation; the label “agent database” alone does not establish that the architecture must be rebuilt. Third, translating natural language into SQL still requires business semantics. He recounts a large bank’s substantial accuracy improvements after years of investment in a semantic and context layer, distinguishing that work from downloading a generic tool. Public benchmarks may have entered training data, so their results do not directly establish business accuracy. Whether information comes from files or a dedicated database does not change the text ultimately supplied to a model; transactional safety and shared state across agents are important reasons to use a database system. Fourth, he believes LLMs can quickly handle many common tuning tasks, while differences between test and production workloads and difficult edge cases still require specialized methods. People must still judge whether the data model represents the business and understand what agents have done. He retains an explicit uncertainty: whether agent workloads require a fundamentally new database architecture remains unanswered.