Supriya Kashyap

@cursor_ai This is pretty cool - Teams integration plans?

Introducing /automate, a skill for agents to set up automations for you. Describe your task in plain language. Cursor configures the triggers, instructions, and tools.

@ClaudeDevs Context sharing is so helpful!

Artifacts are now live in Claude Code. Ask Claude to turn what it's working on into a page and send the link to your team. The page updates as the session keeps working. Available today on Team and Enterprise plans.

@NoamShazeer Half the techniques that make inference affordable today have your fingerprints on them. Congratulations @NoamShazeer, eager to see what comes next!

I’m excited to share that I’ll be joining OpenAI and look forward to working with the exceptional team there. It was a difficult decision to move on. I’m incredibly proud of the amazing team at Google and everything we’ve built together. It has been an honor and a pleasure to work with all of you.

@lennysan Would love to see more representations from neoclouds!

Who else should I have on this year?

Upcoming podcast guests + Jeff Dean, Chief Scientist at Google DeepMind + Andrew Ambrosino, Head of PM and Eng for Codex + Fiona Fung, Head of Eng for Claude Code/Cowork + Tara Seshan, Head of ChatGPT, Productivity at OpenAI + Dianne Penn, Head of Product, Research at Anthropic + Elizabeth Stone, CPTO at Netflix

@Cursor built a $4B ARR by being the best orchestration layer for code: composer, agent loops, multi-file context. But its models were borrowed and now with the $60B @SpaceX acquisition, they break that dependency. Lesson I am taking: Great UX + orchestration can take you surprisingly far on the borrowed inference. But at the frontier, you eventually have to own your compute stack. The merging of Cursor's application layer with the Colossus infrastructure is an example of how the app layer and compute layer are collapsing into each other.

Played around with Claude Code’s Dynamic workflows feature over the weekend for a codebase-scale migration script. My observations: The spin-up latency is real but also watching it orchestrate hundreds of parallel sub-agents to rip through multi-file refactoring without choking the prompt cache was a nice product experience. Learning: Need to practice more on building clear completion conditions so the agent knows exactly when to stop spinning up sub-agents and actually merge the PR, without causing runway token burn.

Agree. The real value lies in the execution architecture built around the foundation models now. As companies build private eval systems, persistent agent environments, and localized context loops, it helps convert the non-deterministic models to stable, proprietary infrastructure. The moat truly lies in owning the learning loop where enterprise workflows dynamically train the agent.

Scanning X this week, I noticed three macro shifts in the Frontier AI space: 1. Persistent Agent Architecture: OpenAI acquiring Ona is a product signal, indicating that the bottleneck for tools (Cursor, Claude) is shifting from model capabilities to giving the agents secure, customer-controlled cloud sandboxes to run week-long code migrations 2. Transaction Layer: Again, OpenAI + Visa indicates agents are getting native financial identities. Fascinating how we've moved from just generating texts to initiating secure commerce within deterministic network tokens and spending limits 3. Compute economy: MAI debuting 7 new models to manage massive compute increases, while NVIDIA locked in a multiyear SKH deal for AI Factory hardware. The frontier is entirely about securing the raw processing and memory footprint needed to sustain heavy agentic loads. Incredible to witness this scale play out in real-time!

From a warehouse in El Segundo to the biggest IPO in history. $SPCX is officially live on the Nasdaq. Exciting to see rockets, Starlink, and space-bound AI all wrapped up into a single $1.7T+ ticker. Inspiring to watch this level of scale and execution!

Thrilled to partner with the Claude team on this! 🤝 Claude Managed Agents just got more powerful with @superserve_ai as their execution environment With full memory snapshots, infinite session lengths, programmable egress controls, and massive parallelization, your agents get to freely explore with you completely in control.