How Vercel AI Gateway keeps Anthropic models online

#1 𝕏

Guillermo Rauch built multiple token origins for Anthropic’s models in Vercel AI Gateway to keep them online nearly 100% of the time, and rolled out routing rules to automatically swap out any retired model.

#2 𝕏

Harrison Chase observes the agent industry pivoting from frameworks like LangChain, AI SDK, and LlamaIndex to full-fledged harnesses such as DeepAgents, Claude Agent SDK, and EVE—with DeepAgents predating EVE by about ten months.

#3 ▶️

OpenAI PM Reveals How He Uses Codex to Do Product Work | Rohan Varma

Peter Yang

Also covered by: @Peter Yang

#4 📝 Jesse Vincent

Some new agentic patterns - Prime Radiant runs agents like Scribble, Nora, Spec-together and a new Sen built on Lace; Sen 2.0’s architecture prevents the main agent from communicating externally, uses ephemeral subagents for external actions, stores all credentials in a 1Password vault, and gates credential use through an arbiter agent that issues random tokens which a transparent MITM proxy rewrites into real credentials. For browser-only logins the subagent fills a temporary random password via superpowers-chrome and a helper tool replaces it with the real password after arbiter approval, and development work happens in a Claude Code session connected to Slack via the Slackline command-line client.

#5 📝 Simon Willison

sqlite-utils 4.0rc2, mostly written by Claude Fable (for about $149.25) - A long-form post describing the author's experiment using Claude Fable to help prepare sqlite-utils 4.0rc2, including motivations, workflow, and outcomes of using LLM assistance.

#6 📝 Surge AI Blog

Antidote: Optimizing for You - Antidote is a new evaluation that measures which answer users would still be pleased with a month later, graded by domain experts, in contrast to LMArena which measures instant preference. It aims to capture long-term usefulness rather than two-second choices.

#7 📝 Surge AI Blog

Deeper Instructions, Stronger Generalization: Training on ComplexConstraints - Surge trained a 4B model on 1,000 expert-written rubrics from the ComplexConstraints benchmark and achieved parity with a model 60x larger. The improvements transferred to external benchmarks the model had never seen, demonstrating that richer instruction data can yield stronger generalization.

#8 in

Siddharth Srinivasan builds on Peter Yang’s insights, questioning whether enterprise, SMB, consumer and developer super-apps will converge or follow the Perplexity/Manus playbook.

Also covered by: @Peter Yang

#9 𝕏

Logan Kilpatrick highlights that builder Ammaar constantly tests every model and just ported a childhood game using Gemini, showcasing frontier capabilities and setting a benchmark for upcoming Gemini versions.

#10 📝 Mario Zechner

Growing on Purpose: The Work That Makes You. Jeremy Howard on human flourishing in the time of AI. - Jeremy Howard — founder of Fastmail, co‑founder of fast.ai and founding CEO of Answer.ai, and author of the ULMFiT technique — cites 30–50 years of research (including a review of hundreds of experiments) to argue that self‑determination theory’s dimensions of autonomy, mastery, relatedness and purpose produce human flourishing, with authentic motivation boosting vitality, creativity and performance rather than just productivity. He also highlights behavioral activation as comparably or more effective than antidepressants for lifting depression, defines flow (illustrated by racing a motorcycle around Phillip Island) and warns of “dark flow,” while focusing his practical recommendations on autonomy and mastery as work changes under AI.