TL;DR
Albert said Claude’s time horizon jumped past rivals, while Askell argued alignment should teach models what to be, not just what to avoid. Levie pushed token budgets over headcount, and Tan said personal software finally turned real.
An early Claude Mythos Preview snapshot reportedly more than doubles the next-best model on METR’s 80% success-rate benchmark. That’s a big signal that Anthropic is pushing hard on long-horizon task execution, not just chat quality.
The more skills you give Codex, the less you have to micromanage it — that’s the core point here. It’s a clean argument for building agents with real capabilities instead of endlessly polishing prompts.
She argues alignment work shouldn’t only be about stopping bad behavior — it should also give models a clear, positive vision of what they’re for and why. That’s a more constructive framing for Anthropic’s safety work: not just constraint, but shaping the model’s character.
As agents take on longer-running work, token budgeting becomes a real enterprise planning problem — more like managing headcount, marketing spend, or lunch budgets than classic IT costs. He says companies will need new software and new controls to route compute to the highest-value work, and that this could spawn a startup category of its own.
He says he’s stopped writing markdown for most things and now uses Claude Code to generate HTML instead. The take: HTML is becoming the better default for AI-assisted docs and notes, because it’s more flexible and easier for models to shape into something useful.
Built a browser extension that watches YouTube with you and answers questions about what was just said using OpenAI’s Realtime API. The neat bit: it separates the video’s audio from your voice, so it stays quiet unless you actually ask something. Also pushed 32 HTML slide templates into AnyGen, making them plug-and-play even without a coding agent.
He argues agent pricing won’t stay purely consumption-based: token costs matter, but enterprise agents also need identities, roles, auth, budgets, and audit logs. That starts to look less like metered API usage and more like a weird new kind of seat model.
He says the next wave is "personal software" — and points to a 1M-token coding agent that supposedly runs on a 128GB MacBook Pro as the proof point. The vibe: AI is pushing serious software creation onto a single machine, which is exactly the kind of shift YC loves to see.
He says the market is still 3–4 months behind the AI curve, which creates a real edge for people willing to act before the crowd catches up. His example: Codex will be mainstream hype soon, so the smart money should already be positioning now.
Claude’s platform is becoming the agent infrastructure layer, not just an API
The Takeaway: Claude’s platform is shifting from “model access” to the full stack needed to ship autonomous agents fast.
Angela, head of product for Claude’s platform at Anthropic, and Caitlin, head of engineering, describe a philosophy that’s more pragmatic than flashy: make the model easier to use by baking in the boring parts. Their view is that the platform should evolve toward “whatever it’s like the set of primitives and infrastructure that enables you to basically get the outcome as fast as possible.” That means messages API, file systems, skills, code execution, web search, memory, and managed infrastructure — not just tokens in and out.
Their sharpest point is that most teams misjudge where the pain lives. People assume harness engineering is the hard part, but the wall usually shows up later: “everyone hits an infrastructure wall.” Once an agent works in a Mac mini or a quick prototype, production becomes a mess of uptime, state, storage, security, and long-running jobs. That’s why Claude Managed Agents exists: Anthropic built the thing it kept rebuilding for itself.
They also argue the old “generic harness, swap models later” mindset is fading. As models diverge, the best results come from pairing the harness and model more deliberately. In other words, the platform is no longer neutral plumbing — it’s part of the model’s behavior, and that path dependence matters.
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.
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.