Follow builders, not influencers.
TL;DR
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.
Salesforce’s Claude Code rollout is the real story: they didn’t just make existing work faster, they deleted steps, removed handoffs, and let agents own chunks end to end. A migration they thought would take 231 days shipped in 13, and one PR landed 21 endpoints with 100% test coverage. Boris says the kicker is quality didn’t slip — incidents fell 5% even as output climbed.
He says founders keep mistaking funding for traction: if people don’t want the product yet, more cash just pours gasoline on a fire that isn’t lit. The sharper YC take here is simple — stop chasing money and go make the first fire.
A company spending $500M to recreate an app layer is basically a giant billboard for software. The nuance matters, but the takeaway is clear: if a business is willing to pay that much to own the experience, the app layer still has plenty of value left.
AI agents need guardrails before they become enterprise liabilities
The Takeaway: The real AI security problem isn’t prompts — it’s autonomous agents taking the wrong actions at scale.
Maxim Bar Kogan, cofounder and CEO of Onyx Security, is building what he calls an “AI control plane” for enterprises: agents that watch other agents. His bet started early, when AutoGPT hinted that LLMs could stop generating text and start taking actions. Back then, most buyers thought he was too early. Now, with coding agents like Claude Code and enterprise adoption accelerating, the risk has caught up to the vision.
His core argument is blunt: traditional security breaks down when software is allowed to act like a human. Identity tools assume permissions can be tightly scoped, but agents need broad access to be useful. Endpoint and API tools can log what happened, but they can’t tell whether the model was justified in doing it. As he puts it, the hard part is “understanding what another AI system is thinking, what is it planning to do.”
Onyx’s answer is not to run a giant model on every action. That would be too slow and too expensive. Instead, they train small models that are good at one job: deciding when a higher-level review is needed. It’s a blitz-chess mindset — spend almost no compute on routine moves, then slow down hard when the position gets dangerous. Kogan thinks that same logic will define AI governance as models get smarter and more autonomous.
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.
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.