AI Builders
Digest

Box CEO Aaron Levie highlights that enterprises are in day one of agentic AI rollout. While coding agents have taken off, diffusion to broader knowledge work faces hurdles like access controls, data integrity, and massive token costs. Breakthroughs outpace implementation, making stable architectures hard. Levie notes a shift from IT budgets to line-of-business OpEx, with needs for new FinOps-like tools for AI compute ROI. Quote: "The breakthroughs keep happening faster than the customer can implement any kind of standard architecture."

Anthropic's Boris Cherny shares five tips for running Claude Opus autonomously for hours/days: auto mode permissions, dynamic workflows, /goal or /loop commands, cloud usage, and self-verification methods. Peter Steinberger emphasizes designing loops that prompt agents instead of manual prompting. Sam Altman notes interesting recursive loops in AI usage incentives.

Vercel CEO Guillermo Rauch announces their AI Gateway recovers on average over 1T tokens a month with zero markup, providing redundancy, observability, and usage controls similar to Stripe's payment optimizations. This helps manage costs in high-volume AI workloads.

Box CEO Aaron Levie details new markdown editor, CLI support, and deep integration with agents like Claude Cowork and Codex. Headless usage will drive consumption models alongside seats. Levie sees dual UI and agentic interfaces persisting.

Claude Blog announces dreaming for self-improving agents via pattern extraction from past sessions, outcomes with rubrics for quality grading, and multiagent orchestration for complex tasks. Harvey, Netflix, and others report significant gains in completion rates and efficiency. Link: https://claude.com/blog/new-in-claude-managed-agents

Thibault Sottiaux from OpenAI introduces a 100-day program selecting one impressive Codex user daily for 10X usage limits to push boundaries. Replit CEO Amjad Masad celebrates builders focusing on market and results.

Former Google PM Madhu Guru counters the misconception that training data is low-skill work, emphasizing complex domain-specific knowledge needed for frontier models beyond SWE. Companies like Mercor are doing high-leverage work.

Y Combinator President Garry Tan highlights education on AI tools as a major bottleneck and shares GBrain updates for tracking thinking evolution. Aditya Agarwal reflects on IPO liquidity fueling new ventures.