TL;DR
Steinberger said AI agents need lean skills, not bloated prose; Levie argued AI won’t kill jobs, it’ll raise the bar. Yang said Codex tested itself while Claude still owned frontend, and Garry Tan pushed hard evals as the way agents got better fast.
He says most agent skills are wasting context on long-winded descriptions, and he built a tool to find the worst offenders. The bigger point: if you want agents to stay fast and cheap, token efficiency has to be part of the spec, not an afterthought. He also shipped a dependency purge in OpenClaw, swapping Sharp and Jimp for a 2MB WebAssembly Rust image pipeline called photon.
He argues the pessimistic AI-and-jobs take misses how work actually changes: automation usually doesn’t shrink demand, it resets expectations upward. As a Box CEO, he points to finance, legal, software, and healthcare getting more analysis, advice, and depth — not less — once the tools get better.
Codex is getting strong marks for browsing and testing its own work, but frontend and design tasks still seem to be Claude’s lane. He also argues the old “ship the MVP first, systems later” advice is flipped in the AI-agent era: documentation, cron jobs, and skill files are now what let small teams move like 10 people.
He says the only way to stay at the frontier is to build, because priors in AI are getting stale every few months. The bigger point: if you’re still doing knowledge work the old way while “alien super intelligence” is here, you’re choosing comfort over learning — automate or get automated.
He says the trick to making agents actually useful is brutally simple: feed them context, run them through multiple frontier-model evals, and keep tightening the skill file until the score improves. Pair that with code, unit tests, and LLM-as-judge checks, and the gains stick instead of evaporating.
He pointed to an internal @every counterpoint to "After Automation," suggesting the real debate isn’t whether AI replaces work, but how humans stay meaningfully in the loop. He also joked that the Pope’s warning about a new Tower of Babel sounds a lot like @every’s 2024 AI framing.
Agents need computers, not just models.
The Takeaway: Agents only become useful when you give them a real computer, not just a chat box.
Ivan Burazin, CEO of Daytona, has spent 16 years building developer tools, and his core thesis is blunt: agents are “digital knowledge workers,” so they need “at least one sandbox, sometimes more.” That insight came from a practical failure. He tried using Claude to fetch bank data and hit a wall when it asked for direct access. “Log in and give me a no. I will not give you access,” he said. That moment convinced him the agent needed its own machine, its own account, even its own phone number for 2FA.
His bigger point is that the infrastructure underneath AI is being rebuilt around stateful work. Traditional hyperscalers were optimized for apps that shouldn’t change on the fly; agents are the opposite. They need persistence, concurrency, and the ability to keep working after your laptop sleeps. That’s why Daytona focuses on the sandbox layer rather than the harness or the model itself.
Burazin also sees a gap in the stack: models don’t really learn on the job yet, memory is still awkwardly handled with markdown files, and most enterprise work still lives inside legacy desktop apps. His view is less “AI will change everything” and more “the computer is the product.”
Peter Steinberger called AI code review the killer dev workflow, while Thariq said Claude Code turned files into a workbench. Cursor bet on specialized models plus brutal RL infra, and Anthropic detailed how it contained Claude across products.
Levie warned AI demos fool CEOs; the hard work starts after. Tan said high agency plus taste beat everything, while Yang urged builders to build the system that builds the MVP first. LeCun kept the pressure on: LLMs help, but they’re not the road to real intelligence.
Levie said AI kills tasks, not jobs, so roles just expand. Masad said Replit’s AI stack can ship MVPs in a weekend, while Tan showed Qwen fine-tunes in hours and personal AI got real. No Priors said Cerebras won by betting on speed before the market cared.
Levie said AI will surface more bugs, making humans the bottleneck. Steinberger argued GitHub’s native PR limits kill bot policing, while Tan pushed the “ship the 60%” rule and Swyx said AI coding should harden codebases, not just move faster.
Altman shipped Codex while asking what AI should fix. Levie saw AI pricing split into tiers, Masad pushed buy-first rewards, and Anthropic added Auto Mode to Claude Code as the reliability bar jumped.