TL;DR
Rauch said the moat moved from code to the code factory, while Levie argued every team now needed an agent wrangler. Cursor leaned into customizable multi-agent views, Replit added region controls, and No Priors backed Periodic Labs’ bet that AI could learn atoms by running experiments.
He says off-the-shelf coding agents break down in big monorepos, so companies are building their own AI software factories with custom knowledge, workflows, and integrations. Vercel just open sourced Open Agents, a reference platform for cloud coding agents, and he’s betting the real advantage is now the means of production — not the code itself.
He says enterprises are about to create a new role: the agent deployer/manager, someone who finds the highest-leverage workflows and wires agents into them. Think 100x faster lead triage, contract review, onboarding, and knowledge ops — with the hard part being context, evals, human handoffs, and ongoing KPI management. As Box’s CEO, he’s basically arguing this becomes a core operating function, not a centralized AI side quest.
He says ~80% of the world’s agent and AI engineering happens in just three square miles — a blunt reminder that the ecosystem is still highly concentrated. He also notes Cognition usage has roughly doubled globally after two launches, with people getting creative once they can compose agents and make them proactive.
Software engineering in 2026, he says, needs two roles: the pirate who hacks fast to find value, and the architect who turns the mess into something durable. As CEO of Every, he’s basically arguing that AI speeds up exploration, but the real edge is still in cleaning up and systematizing what works.
They teased more ways to split the workspace up, down, left, and right — plus more customizations and multi-agent views coming to Cursor. It reads like a push to make the editor feel less like one chat box and more like a configurable command center for serious AI coding.
You can now configure app hosting region in Replit, a practical move for teams dealing with privacy rules and data residency. It’s a small feature with big enterprise implications: less friction for regulated customers, more reason to trust the platform.
He says frontier labs have a sneaky advantage: truly unlimited tokens. In his telling, people join a hot agentic startup, hit token caps and cost constraints, then bounce back to the lab where they can just build without watching every inference bill.
He says GBrain is basically his own OpenClaw/Hermes agent setup, now with opinionated search, skill packs, and a voice agent built on OpenAI Realtime — with Gemini Live next. He also shipped v0.9.3, adding search tuning, evals, CJK query support, better health checks, and security hotfixes.
He says the real question with OpenClaw, Claude Code, and similar tools is whether they’re actually getting work done — or just making the setup itself feel productive. He also argues OpenAI has a problem if GPT integration isn’t as good as Opus inside OpenClaw, because that’s now the baseline users expect.
Periodic Labs wants AI to learn atoms by running experiments
The Takeaway: AI gets truly useful in science only when it stops guessing from text and starts learning from experiments.
Liam Fedus, co-founder of Periodic Labs and one of the creators of ChatGPT, keeps circling back to the same idea: language models were the start, not the destination. His path runs from physics and dark matter research to Google Brain, then OpenAI, where he worked on productionizing GPT-4 and helped build ChatGPT. That background matters, because his view of AI is unusually grounded: the next leap won’t come from bigger chatbots alone, but from systems that can touch reality.
His core argument is blunt: “Science ultimately isn’t sitting in a room thinking really hard. You have to conduct experiments.” Periodic is built around that premise. Instead of treating data as a static corpus, it uses experimental results as part of an active loop—spotting anomalies, comparing against simulations and literature, then choosing the next experiments. That’s a very different game from training on the internet.
Fedus is also skeptical of the idea that intelligence just scales smoothly. In his view, these systems are “spiky”: world-class in one domain, surprisingly weak in another. That’s why Periodic leans on language models as an orchestration layer, while specialized neural nets handle atomic systems, symmetry, and control. The bigger vision is simple and ambitious: give humanity “agency for atomic rearrangement synthesis” and speed up the physical world the way software has already accelerated the digital one.
Boris Cherny said agentic workflows beat speedups by a mile. Garry Tan argued money amplifies demand, not creates it. Aaron Levie called a $500M custom build a bullish ad for software, while No Priors warned agents need guardrails before they become enterprise liabilities.
Matt Turck said enterprise AI had moved from chatbots to messy, expensive agents, and the bill was coming due. He framed the shift as real progress with real pain: more autonomy, more complexity, more cost.
Every argued AI lowered the floor but raised demand for experts. Anthropic fixed three Claude Code regressions, reset limits, and shipped Managed Agents to split agent brains from tools. Claude also added self-hosted sandboxes and private MCP tunnels.
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.
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.