WorkOS

More than 70% of Neo4j's new business last quarter came from AI use cases, not traditional graph database queries. CTO Philip Rathle explained why on the ground at AI Engineer World's Fair: LLMs are the "right brain," but they need a deterministic, explainable "left brain" for the small set of questions where 100% accuracy actually matters (fraud and compliance). His framing: dropping a brilliant model into your company is like hiring the smartest person you know, but without your company's data and context, that intelligence is meaningless.

Thank you to everyone who came out last night for Dwarkesh Unplugged. And special thanks to Dr. Fei-Fei Li and Dwarkesh Patel for an unforgettable conversation. ✨

For anyone building, evaluating, or trying to understand the infrastructure behind AI agents: join us for a live demo night in SF. Two focused demos on the identity and memory agents need for production: auth.md, WorkOS's open-source agent authentication protocol, and Pieces local-first workstream memory. Plus food, drinks, and time to meet other builders. ⏰ 5:30p PT 📅 July 30 📍 WorkOS SF HQ Register here: https://luma.com/715ul99e

20+ live demos at our office last night. Five minutes each, no slides, no pitching. Just show the thing you built. A few that stuck with us: • Rhys Sullivan (Executor) → had an agent write its own lightweight TypeScript integrations straight from an OpenAPI spec, instead of hand-rolling MCP. • Akhilesh Rangani (usecharming) → demoed agents that build and then use their own apps. • Alex Boone (Neutral Base LLC) → made an app agent-native in under five minutes: zero refactoring, no DOM parsing. • Connor Paton (Tembo) → made the case for moving 100% of dev work to the cloud, on sandboxes that snapshot up to 128GB of RAM. • Kyle Mistele (HumanLayer) → watch and steer a teammate's coding agent live, with a real-time diff viewer instead of waiting on a PR. • Russell Haering (C1) → WASM DB: an object-store database you query in JavaScript that generates its own UI. Basically headless Notion. • Brian Douglas (Paper Compute Company) → tapes, high-fidelity agent traces, powering a self-healing Pokémon Red speedrun. • Nikhil Tayal (NoInfra) → a personal agent anyone can deploy, built first for his mom, down to its own inference gateway and wallet. • Matt Dailey (Ref) → moved collaboration to the decision layer: a shared spec doc you launch agents from before anything gets built. • Mike Clarke (Modem) → shipped two open-source tools: HUNK, a GitHub-style review UI in your terminal, and Sideshow, which turns a 100-file agent branch into a shareable explainer. • Michael Grinich (WorkOS) → a LinkedIn client he vibe-coded over a weekend. 15,000 lines of TypeScript, ~900 tests, a local sync engine, and not one line he read himself. Claude wrote all of it. And more! The through-line: real software shipped at a speed that didn't exist a year ago. Full recording👇

Tomorrow! Two of the most interesting people in AI. One stage. One night in SF. Dwarkesh Patel interviews Dr. Fei-Fei Li on the forces shaping technology and AI — and what it means for the future. Plus, Michael Grinich sits down with Dwarkesh for a candid conversation on how he got his start. Link👇

Demo Night @ Work OS Whether it's a side project, an internal tool, an open source project, or a startup, this is your chance to show what you've been working on and see what others have created. The only rules: Live demos only No slides No company pitches Come share your work, get feedback, and meet fellow builders in the community.

Enterprise Ready Conf returns to SFJAZZ on October 7. Speakers announced soon. Register👇

In a Nix-based architecture, every dependency is pinned to a file path. So a known CVE already has a known address. You look it up, see exactly where it lives, and remediate by bumping the version up or down to the last one that worked. Ron Efroni, CEO of Flox and President of the NixOS Foundation, calls this the "cryptographic layer" of software: deterministic, reproducible, and hardened by construction. We covered this and how Flox runs agents and 100+ sandbox platforms inside portable, hermetic environments.

Speed is the next unlock for browser agents. Paul Klein IV, founder and CEO of Browserbase, predicts that within six months computer-use models will automate web pages in real time: as fast as a human, then faster. Accuracy is largely handled now. Model quality is up and the harnesses around agents have matured, so reliable agents in production are already real. The remaining gap is watching an agent slowly click through a page you could have finished in seconds. Michael Grinich sat down with Paul to discus the state of browser agents, what Browserbase is shipping, and why speed is the benchmark that proves automation has arrived.