AI Builders
Digest

Anthropic Head of Product for Cloud Platform Angela and Head of Engineering Caitlin discussed the evolution of the platform from basic completion APIs to sophisticated managed agents with memory and infrastructure. The key philosophy is building higher-order abstractions so users get the best outcomes from Claude with minimal work. In the future, Claude will understand itself well enough to spin up sub-agents on the fly; users will primarily specify outcomes and budgets. Scaling the platform for constantly running, self-modifying agents is a major focus.

Box CEO Aaron Levie observes that as agents tackle more long-running tasks consuming vast compute, token budgeting has become a major topic in large enterprises. Companies will manage tokens as carefully as talent or marketing spend to avoid budget blowouts on low-value work. This creates opportunities for new software and shifts agentic spend into organizational budgets.

Anthropic's Thariq (Claude Code) shared that he has stopped writing markdown files for almost everything and now uses Claude Code to generate HTML instead. This offers superior presentation and interactivity. He shared example HTML documents.

Peter Steinberger (Polyagentmorous ClawFather at OpenClaw + OpenAI) noted: "The more skills you give codex, the less you have to prompt." This improves developer productivity significantly.

Zara Zhang built a "YouTube realtime copilot" browser extension with OpenAI's realtime 2 API. The agent watches videos alongside you and answers questions via realtime voice chat. It differentiates the video's audio stream from your voice and stays silent unless asked.

YC President & CEO Garry Tan expressed excitement about 1M token context windows with usable coding agents running on a 128GB MacBook Pro, declaring "Personal software is coming."

Matt Turck (FirstMark) noted that pure consumption-based pricing for AI agents is becoming less obvious. Enterprise agents need identities, roles, auth, budgets, and audit logs, sounding more "seat-like" (but not human seats).

Anthropic philosopher and ethicist Amanda Askell highlighted that alignment research should focus not only on averting concerning behaviors but on giving models an honest and positive vision for what AI can be. She is excited about the future of this work.

Anthropic researcher Alex Albert shared that an early Claude Mythos Preview snapshot provided to METR has a time horizon more than 2x the next best model on their 80% success rate benchmark.

Every CEO Dan Shipper noted a generational opportunity in AI due to time lags. People at Milken are 3-4 months behind; investing now in areas like Codex will pay off as excitement builds in the coming months.