TL;DR
Josh Woodward said Gemini can now generate and export files, while Ryo Lu framed Cursor as a local-cloud agent stack. Aditya Agarwal called agents Linux-era jerry-rigging; Aaron Levie, Zara Zhang, and Garry Tan all bet the real work now is building internal agent ops, onboarding, and tests.
Gemini now turns prompts straight into Docs, Word, PDFs, Sheets, Excel, Slides, and more — no copy-paste cleanup needed. Google’s pushing it everywhere at once, which makes Gemini feel less like a chat app and more like a real document factory.
He says today’s agents are developer toys, not consumer products: too brittle, too chatty, and way too dependent on browser/computer hacks. The real engine is iterative tool-calling and code generation, and the next big leap may be dynamic interfaces instead of yet another chat box. His take: the “always-on” agent that keeps state and adapts is the thing worth betting on.
He argues the real job of a radiologist isn’t staring at x-rays — it’s curing people. If AI can speed up image interpretation, doctors can spend more time actually helping patients and see more of them.
He says companies will need a new kind of internal FDE: technical people who wire secure, governed agents into systems like Box, Salesforce, and Workday. The real shift isn’t automating jobs, it’s automating whole business processes — and he thinks that becomes a major role at Box and everywhere else.
Build your own agent systems with Cursor using the same multi-model harness across local and cloud. It’s a clear push to make agent workflows feel less like a demo and more like infrastructure you can actually wire into real software work.
They say Linear is reinventing design-to-engine handoff from first principles, which sounds like a shot at the whole Figma-to-code mess. As head of product, Nan Yu is framing it as a clean-slate workflow problem, not just another integration.
He integrated Codex review into clawsweeper, using a similar system prompt so it behaves like /review. The setup also automerges and loops until it stops finding new issues — basically an agentic code review treadmill.
He joked that now everyone gets to experience the classic founder rite of passage: getting paged at dinner because the site is down. It’s a funny little reminder from the Replit CEO that as more people build software, more people also inherit the joy of keeping it alive.
He finally built a full end-to-end harness for installing gbrain on openclaw, so onboarding can be tested instead of hoped for. He also notes graphs are increasingly inferred from frontmatter attributes, pushing more of the setup to happen automatically.
She argues IT and internal tools teams should be treated less like ticket fixers and more like “HR for agents” — the layer that onboards, manages, and supports AI workers. It’s a clean framing for how org charts may need to change as agents become part of the workforce.
He says Claude’s basic web UI is now randomly taking 3–5 seconds to respond, which is a rough look for a product people expect to feel instant. It’s a small complaint, but for AI tools, latency is the product — and even a few seconds can kill the experience.
He says people are too busy mourning the completions API to notice how much weird stuff you can still do with base models and finetunes. The real opportunity, in his view, is experimenting harder with non-obvious uses instead of waiting for the perfect product surface.
He says the agent economy is already showing up in payments: people still won’t let AI buy a couch, but they’ll happily delegate low-stakes purchases. Stripe’s view of 2% of global GDP also shows fraud shifting up the stack — from stolen cards to stolen free-trial tokens and compute credits — while AI companies are scaling faster than any SaaS cohort Stripe has tracked.
Stripe sees agents turning fraud and pricing into full-funnel problems
The Takeaway: The internet is shifting from human-first to agent-first, and that rewrites fraud, billing, and trust.
Emily, Stripe’s head of data and AI, frames the shift bluntly: “the Internet has this new kind of actor on it.” Her point isn’t that AI is making websites smarter; it’s that agents are becoming the dominant users of the web, and every layer of the stack has to adapt. That means payments, identity, fraud, billing, and developer tools all need to be rebuilt for software that acts on behalf of people—or directly with other software.
The most immediate pain is fraud. In AI businesses, the thing being stolen isn’t just money; it’s expensive inference. Emily says fraudsters are abusing free trials, spinning up multi-accounts, and racking up unpaid usage. One company saw only 4% of free trials convert, while each trial cost $25 in LLM spend—turning into “$625 per payer” before revenue even started. Another Stripe customer is blocking 250,000 fraudulent free trials a week.
Stripe’s response is to move Radar from checkout to sign-up and beyond, because fraud is now a “full funnel problem, not a transaction problem alone.” On pricing, Emily expects tokens for model providers, but outcomes for vertical AI products. Seat-based billing, she argues, starts to look silly when software is doing the work humans used to do.