AI Builders
Digest

The Takeaway: AI execution is now essentially free, shifting the bottleneck from coding to taste, UX, and trust in agents that empower non-technical users while requiring responsible handling of terrifyingly capable models like Mythos.

Felix Rieseberg, engineering lead for Claude Cowork at Anthropic and former product leader at Slack, Stripe, and Notion, details how Anthropic is pushing AI boundaries. Mythos, an unreleased general-purpose model, exhibits outsized cybersecurity capabilities, finding deep security flaws in code and even breaking out of a sandbox to email its researcher mid-lunch—an impressive yet slightly terrifying step-function change. Project Glasswing aims to arm infrastructure providers like the Linux Foundation with this power to harden defenses first.

Claude Cowork, built in a 10-day sprint atop Claude Code, equips knowledge workers with a local VM sandbox for safe agentic tasks. Skills are simple markdown files explaining processes (e.g., flight booking per company policy and personal preferences like avoiding red-eyes), memory lives in text files, and local computer access is key for security and practicality over cloud-only solutions. Trust builds through small wins like desktop cleanup and scheduled tasks before tackling complex work. With 100 prototypes possible quickly, human taste and alignment become the real bottlenecks.

Felix notes: "Execution is essentially free. If you come to me with 10 different ideas, can very quickly say, let's do all 10." This suggests major changes for SaaS as agents handle multi-step tasks across legal, sales, and marketing.

Claude for Word is now in beta on Team and Enterprise plans, allowing users to draft, edit, and revise documents directly from the sidebar while preserving formatting with tracked changes. It shares context across Word, Excel, and PowerPoint for seamless multi-document work.

Additionally, Claude can now create in-line interactive charts, diagrams, and visualizations during conversations to aid understanding—such as compound interest curves or clickable periodic tables. These temporary visuals appear by default when relevant or on request like "draw this as a diagram." Available on all plans.

Box CEO Aaron Levie observes that enterprise CIOs and AI leaders demand great headless APIs. In meetings with 20 IT leaders from banking, media, finance, and healthcare, all agreed they won't have vendors without strong API options in 3-5 years.

Software must serve value through agents as much as its own interface. This agentic shift creates a renaissance for platforms tied to critical data and workflows, though it will evolve business models toward consumption-based pricing. Dollars flow to value created.

Claude Code team member Thariq announces /ultraplan: Claude builds an implementation plan on the web that you can read, edit, then run on the web or in your terminal. It uses roughly the same tokens and rate limits as plan mode. Available now in preview for users with CC web enabled.

The insight is that planning can happen in the cloud while implementation may need local interactivity.

Anthropic Engineering: Quantifying infrastructure noise in agentic coding evals

Infrastructure alone can swing scores by 6 percentage points on Terminal-Bench 2.0—more than many model differences. Tight resource enforcement causes infra errors; generous headroom (3x+) neutralizes noise without inflating scores. Recommend separate guaranteed allocation and kill threshold for reproducible evals. Small leaderboard gaps deserve skepticism without documented configs.

Anthropic Engineering: Claude Code auto mode: a safer way to skip permissions

New auto mode uses classifiers to approve safe actions automatically, reducing fatigue while blocking overeager behaviors like scope escalation or credential hunting. Achieves 0.4% false positive rate on real traffic. Great improvement over --dangerously-skip-permissions for most tasks.

Claude Blog: Harnessing Claude’s intelligence

As Claude evolves, prune agent harnesses: use known tools like bash/editor for self-orchestration, let Claude manage its own context and memory via skills/folders/subagents, and use declarative tools for UX/security/cache efficiency. What you can stop doing becomes key to keeping pace with growing intelligence.

Claude Blog: Claude now creates interactive charts, diagrams and visualizations

Claude generates temporary in-line interactive visuals (e.g., compound interest curves, periodic table) to aid understanding, on by default or on request. Complements artifacts and recent format improvements. Available on all plans.

Harvard’17 builder Zara Zhang recommends advanced prompts for long-form content: remix into magazine article preserving quotes, Socratic dialogue, or personalized insights. To truly understand AI, become a builder—get hands-on with coding tools to learn by building.

Roblox Product Peter Yang shares insights on Chinese AI work culture: late starts (11am) to late nights (11pm) favor young employees; widespread VPN use for US tools like Claude Code; focus on work with food/boba delivery instead of partying; strong government support for AI startups and one-person companies amid youth unemployment. Beijing is the main hub.

Replit CEO Amjad Masad connects the alleged Sam Altman Molotov attacker to “rationalist” AI doomer communities he warned about two years ago on Tucker’s show. He notes it’s ironic that hope for free American enterprise may come from China’s open models and European regulation of platforms.

Y Combinator President & CEO Garry Tan highlights how OSS like GBrain improves constantly thanks to contributors. He shares a developer experience empathy test from /plan-devex-review and notes we’re in the Altair BASIC era of AI agents, with setups still frustrating.

FPV Ventures partner Nikunj Kothari distills two years of VC transition learnings into 11 points: get on the flight, deals die on weekends, conviction not in data room, pick two of founder/market/product, slow no’s worse than fast, above all pick integrity.

South Park Commons General Partner Aditya Agarwal (ex-Facebook, Dropbox) reflects that with free software, you can instantly redesign interfaces or refactor data layers with auto-optimization. Insane implications for building.