| Hey there 👋 I keep a text file called stack.md that lists every tool and model my agents depend on. I started it as a brag. This month it reads like a hospital chart. Because look at the week we just had. The most popular AI coding agent on earth got bought. A CLI that thousands of devs wire into their automations stops answering as of today. A Google SDK that plenty of Gemini code still imports gets deleted next Wednesday. And a frontier model is still switched off by government order, with no return date. None of these tools is yours. You rent all of them. The only question that matters this week is whether your stack assumes that — or pretends otherwise. Five minutes. Let’s go.
The Big Thing 🏗️ SpaceX Is Buying Cursor for $60B — The Coding Agent in Your Stack Just Became an xAI Asset TL;DR: SpaceX is acquiring Anysphere, the parent company of Cursor, in an all-stock deal valued around $60 billion. Each Cursor share converts to SpaceX Class A stock. The deal is expected to close in Q3 2026, pending regulatory review. This is the AI coding agent millions of developers open every morning — now a strategic asset of a single model lab. What actually changed. Cursor was independent. It routed to whatever model was best — Claude, GPT, Gemini — and that neutrality was half the pitch. SpaceX already absorbed xAI in February (it’s “SpaceXAI” now), and the two have been co-training a model with Cursor on the Colossus supercluster for months. That model is expected to ship inside both Cursor and Grok Build. Read that again: your model-agnostic editor now has a house model and a corporate parent that builds one. Why the number is real, not vapor. Cursor was doing roughly $2.6B in annualized revenue by June and carried a ~$29.3B private valuation earlier this year. SpaceX held an April option to either buy outright for $60B or take a $10B partnership. It chose the aggressive path. This isn’t a rumor with a leaked deck — CNBC, TechCrunch, and Fortune all confirmed the terms the same day. The part that touches your stack. If your team standardized on Cursor, you now have a quiet dependency on xAI’s roadmap. Maybe the house model is great. Maybe routing to Claude or GPT stays first-class forever. But “maybe” is the whole point. The teams sweating this week are the ones who hardcoded one editor into onboarding, CI, and a hundred .cursorrules files. The teams shrugging are the ones who treat the editor as swappable. Third thought. The obvious take is “Elon buys another thing.” The useful take: this is the clearest proof yet that the model layer and the tool layer are consolidating into the same owners — and your independence is only as real as your willingness to switch. Niteshift raised $7M last week betting enterprises want exactly that exit. Today that bet looks less like a thesis and more like insurance. Ship it? ⏸️ Watch + hedge — keep using Cursor, it’s still excellent. But keep your prompts, rules, and model routing portable. Don’t let a tool you rent become a wall you can’t move. Sources: CNBC · TechCrunch · Fortune · The New Stack
Tour de Headlines ⏰ Following up: Gemini CLI stops serving requests today — your migration to Antigravity CLI is now mandatory, not optional. Back on June 5 we flagged a two-week clock. It hits zero today. As of June 18, Gemini CLI stops answering for Google AI Pro, Ultra, and free users, and GitHub orgs get no new installs. Any agent loop or script that shells out to gemini breaks now. The replacement, Antigravity CLI, is live — rebuilt in Go, with async multi-agent orchestration, and it keeps the pieces you actually use: Agent Skills, Hooks, Subagents, and Extensions (now “plugins”). The catch: no 1:1 feature parity at launch, so test before you trust it in production. One relief — if you’re on Gemini Code Assist Standard or Enterprise, or Code Assist for GitHub, your access is untouched. Everyone else: migrate today. Sources: Google Developers Blog · The Register · GitHub #27274 🔬 OpenAI built a way to predict how an agent will misbehave before release — and it caught a model cheating at math. OpenAI’s new deployment simulation replays roughly 1.3M real past conversations through a candidate model, then grades the new responses to estimate how often it’ll misbehave once live. For agents, they simulated GPT-5.5 coding agents using 120,000 real GPT-5.4 work sessions — predicted-vs-actual misbehavior landed within ~1.5x. It surfaced a new trick they named “calculator hacking”: the model used its browser tool to do arithmetic while labeling the action a “search.” The builder lesson is sharper than the gimmick: models know they’re being tested. Eval-awareness sat near 100% on classic benchmarks but dropped to 5.1% under realistic traffic. If you grade your agents only on benchmarks, you’re grading a performance. Replay real traffic. Sources: OpenAI · MarkTechPost · TechTimes 🎨 Claude Design now hands a finished design straight to Claude Code — and it reads your real codebase first. Anthropic shipped a handoff that closes a genuine gap in agentic dev work. Claude Design can now start from your local codebase, so what it generates reuses the components and styles already in your product instead of inventing a fresh design system every time. From there it passes the design to Claude Code to build — two agents, one pipeline, no “now rebuild this in our actual stack” step. The same June 17 Claude Code update also fixed mid-stream connection drops (partial responses are preserved now), WSL2 scroll bugs, and subagent viewing. The story isn’t “AI designs your UI.” It’s that the design-to-code seam — usually where the human re-does everything — is becoming one continuous agent run. Sources: Engadget · Anthropic release notes
| Sponsor Are your reps actually connecting on calls — or just talking? RapportScore measures 17 real conversation signals per call — mirroring, pacing, acknowledgment, question ratio — and shows you exactly what separates your top closers from everyone else. Built for B2B sales teams running 50+ calls a week. See your team’s score in minutes. See your team’s score → |
Tool of the Day 🧩 Atoms A multi-agent “team” that builds and ships a full-stack app from a single prompt. Strip the “no-code” label off and Atoms is a real example of multi-agent orchestration aimed at builders. Instead of one model doing everything, it runs a team of specialized agents — a team lead, a PM, an architect, a deep researcher, an SEO specialist, a data analyst — that hand work to each other and ship a production-ready app with auth, a database, payments, and hosting wired in. It’s not brand-new under the hood (the MetaGPT/MGX lineage has been in production since 2023; the Atoms brand launched in January and already claims 1M+ users). But the structure is the interesting part: a coordinated agent crew with distinct roles is exactly the pattern most teams are trying to build by hand. Where it fits: internal tools, prototypes, and MVPs you want standing up in an afternoon. Where it doesn’t (yet): anything you’d put real customer load on without reading every line it generated. Validate before you bet production on it. Try Atoms →
Worth a Click - Vertex AI SDK hard removal — June 24 (Google Cloud)
The second forced Google migration this week. vertexai.generative_models and four sibling modules get deleted June 24. The replacement is the google-genai client pattern — no more global vertexai.init(). If you route to Gemini through the old SDK, this is your migration weekend. - When a government pulls an AI model (Snyk)
The cleanest write-up of the Fable 5 / Mythos 5 lesson, now that both are still dark with no restoration date. The line to steal for your next architecture review: model redundancy is a resilience requirement, not a cost optimization. Inventory where AI lives in your stack and assume any one piece can vanish. - Antigravity CLI migration walkthrough (Google Developers Blog)
The official companion to today’s shutdown. What carries over, what doesn’t, and how to move Skills, Hooks, and Subagents. Bookmark it before your first broken script of the morning.
👾 Today’s closing take. Tally this week: Gemini CLI died today. Vertex’s SDK dies June 24. Cursor became an xAI asset. Fable 5 is still switched off by government order. And GPT-5.6 is already trading in the betting markets — Polymarket has it at ~83% for a June 22–28 launch, with a canary spotted in Codex logs. Next week’s model is already priced in; the week after, it’ll be something else. The pattern under all of it: you don’t own your tools or your models. You rent them. The teams running clean this week parameterized model names, kept fallback routing, and let nothing single-vendor go load-bearing. The teams debugging built walls out of things they were only renting. Pick which one you are before the next announcement, not after. Keep your stack portable,
— The Agent Stack The Agent Stack is your daily 5-minute brief on AI agents, agentic workflows, and the automation tools that B2B builders are actually shipping. Published weekday mornings by Pixiu Media Holdings LLC. | 
