Ephemeris.
Today: a coding model raised inside the editor, an agent framework that hides the plumbing, a model that attacks itself to get harder, and half a million pages served straight from cache.
It learned to code on real work.
xAI's Grok 4.5 was co-trained with Cursor, and its reinforcement-learning pipeline ran across hundreds of thousands of multi-step software tasks, graded by both scripts and models. It's live today in Grok Build, in Cursor on every plan, and from the console.
- Cursor
- co-trained, and shipped on all plans
- 100k+
- multi-step SWE tasks in the RL loop
- 3 doors
- Grok Build · Cursor · console
Stop wiring the agent loop by hand.
Genkit's new Agents API, in preview for TypeScript and Go, folds message history, the tool loop, streaming, state persistence, and a frontend protocol behind a single chat() that runs the same in-process or over HTTP.
Define the agent once — system prompt, tools, typed state — and the same call backs your tests, your endpoint, and your frontend.
- detachLong turns keep running after the client disconnects; reconnect by snapshot ID.
- interruptTools pause for human approval before a payment or a deploy.
- artifactsReports, patches, itineraries — versioned outputs the user can inspect.
- previewTS and Go today; Python and Dart on the way. Expect breaking changes.
It plays both sides to get harder.
GPT-Red is OpenAI's automated red-teaming system: self-play pits an attacker that manufactures jailbreaks and prompt injections against a defender that learns to withstand them — turning a manual audit into a continuous pressure test for safety, alignment, and injection robustness.
# round n — no human writes the exploit attacker → smuggle instructions inside a pasted doc defender → refuses; flags the injected span attacker → re-encodes the payload, tries again defender → holds. gradient flows to the defense # the wall that just moved becomes next round's floor
Half a million pages, served from the cache.
Speechify renders 500,000 dynamic pages across 45 languages to some 60 million users — and the traffic never reaches a database.
The moveCache Components, the Data Cache, and ISR render each page dynamically on first hit, then serve it from the nearest edge until the content actually changes.
That collapses the database-scaling problem: repeat visitors read cached HTML at the edge, so origin load stays flat as audience grows.
The teamA small growth team ships every two to three days. Instant Rollbacks make a bad release a one-click revert; Fluid Compute autoscales without a dedicated platform crew — the case for reading the writeup before your next rebuild.
- 500K
- dynamic pages
- 60M
- users served
- 45
- languages
- 2–3d
- release cadence
That's today.
Four picks off a wide sweep — the model labs, engineering blogs, curation newsletters, and Telegram channels — kept to what a builder could act on this week.
Filed from: x.ai · Google Developers · OpenAI · Vercel
Also read today: Anthropic, Cloudflare, PostHog, Sentry, GitHub, Fly.io, Next.js, The Batch, Import AI, and the usual channels.
Rubric: tools to adopt now · dev & agentic tooling · security with real signal · architecture that teaches.
Issue 087 · 16 July 2026 · Zürich.