He says agentic coding is a huge boost for developers, IT teams, and domain experts who want to automate workflows or wire systems together. But it’s not a free pass for everyone to casually build and maintain complex software — the real upside is 100X more software, not everyone becoming a full-time builder.

He’s wiring Codex into every commit on main so it hunts regressions and security issues, then spins up more agents to fix and re-check the fix — up to five loops deep. It’s already caught one of his own bugs in the first 10 minutes, which is a pretty strong demo of agentic CI actually doing useful work.

He says the next software wave is "Codex-native" and friends: apps designed to live inside an agent’s in-app browser, where both the human and the model share full context. His example is PostHog inside Codex, where the agent can write queries, inspect results, and even kick off PRs or DB requests — way tighter than an agent trapped in your browser.

She argues the better move is to have AI generate SVGs so illustrations blend cleanly into product design, especially for HTML slides. She’s already using QuiverAI inside AnyGenIO’s Frontend Slides feature, which makes the point feel less like theory and more like a workflow shift.

He says solo AI builders now have the world as their playground, and that’s the real shift. He also floated a fun benchmark idea for models: judge them by which game era they can one-shot or build through in 30 minutes — and thinks we’re still in the NES era.

He says the team is hunting down the worst Claude Code bugs, from hangs during big file writes to other “white whales.” They’re also making the no-flicker renderer good enough to become the default, so expect a steadier editor experience soon.