Presented by · your advertiser here Reach 50k+ builders who ship AI — this is your slot. Primary sponsor placement, first scroll-stop. Book it » (demo slot) |
Yesterday the news was about what AI can build. Today it's about what you don't own.
A 100K-star open-source tool gets a two-week shutdown notice. ChatGPT starts rewriting your memories without asking. A startup ships a memory layer you carry between vendors. The through-line: the agent stack is sorting itself into the parts you rent and the parts you keep. Knowing the difference is now a planning skill, not a philosophy debate.
Five-minute read. Let's get into it.
The Big Thing
Google is shutting the open-source Gemini CLI on June 18. Plan your move now.
Verdict: Watch (with a hard clock).
Here's the situation. On June 18, 2026, Google stops serving requests to the open-source Gemini CLI and the Gemini Code Assist IDE extensions for everyone on the free, Pro, and Ultra tiers, plus individual Code Assist subscribers. The replacement is the Antigravity CLI (`agy`), Google's closed, agent-first terminal. We confirmed the date across Google's own developer blog, The Register, and the Hacker News thread — it's real, and it's two weeks out.
The part builders are reacting to isn't the deprecation itself. It's the shape of it. The Gemini CLI shipped as open source and pulled in 6,000+ community pull requests on its way to 100K stars. That community helped harden a tool whose future is now gated: enterprise Code Assist Standard and Enterprise licenses keep access indefinitely, while the people who filed the PRs get moved to a closed product. That's the "trust crisis" framing in the early coverage, and it's worth naming plainly without turning it into a Google-specific morality play. Every vendor that open-sources a developer tool is making an implicit promise about its lifespan. This is a reminder to read that promise carefully before you standardize on it.
The economics matter more than the optics. Early Antigravity CLI users report burning through weekly quotas in under 2,000 lines of generated code. If your team runs agentic dev loops — where a single task can churn thousands of lines across retries — that ceiling arrives fast. The free-tier math on Antigravity looks worse than what you have today, and the generous Gemini CLI quotas were part of why teams adopted it in the first place.
The migration itself is not the hard part. You install `agy` from antigravity.google, run `agy plugin import gemini` to bring over extensions, rename `GEMINI.md` to `AGENTS.md`, move `.gemini/skills/` to `.agents/skills/`, and swap `gemini` for `agy` in your CI/CD scripts. Google says Agent Skills, Hooks, Subagents, and Extensions all carry over, though not at full feature parity on day one.
Ship it? This is a decision, not a download. If Gemini CLI is load-bearing in your pipeline, you have two weeks to either migrate to `agy` and re-benchmark the new quotas against your real workload, or treat June 18 as the moment to revisit whether a single-vendor CLI belongs at the center of your stack at all. Open alternatives exist; so does staying put and eating the closed economics. What you can't do is wait. Pick the lane this week, test it next week, and don't let a Tuesday shutdown make the choice for you.
Source: Google Developers Blog · vpsmac analysis · The Register
Tour de Headlines
🧠 OpenAI rebuilds ChatGPT memory with "Dreaming V3"
OpenAI replaced the manual saved-memories list with a background process it calls Dreaming V3. Instead of you tapping "remember this," the model synthesizes context from conversations on its own and rewrites old memories as facts change — "going to Singapore in July" quietly becomes "went to Singapore in July 2026." OpenAI reports factual recall jumping from 41.5% to 82.8% on its internal eval, with a roughly 5x compute cut that makes a free-tier rollout viable. For agent builders, self-updating persistent memory is a foundational primitive — it changes how much state you have to manage externally, and how much you trust a closed platform to hold. Plus and Pro in the US first; Free, Go, and more countries to follow.
Source: OpenAI
🛍️ Meta takes its Business Agent global and ships an enterprise platform
Meta moved its long-tested Business Agent to global GA across WhatsApp, Instagram, and Messenger, and unveiled a separate Business Agent Platform for large orgs. The platform lets enterprises configure their own agents and wire them into Shopify, Zendesk, and Shopee, with token-based pricing on the way. If you build conversational-commerce or support agents, this is a major new distribution surface — billions of inboxes you didn't have to win on your own. It's also a competitive marker for the B2B agent vendors whose whole business lives on messaging. Worth watching what the token pricing actually lands at, because that's what decides whether building on Meta's rails beats running your own.
Source: TechCrunch
🧩 Nemotron 3 Ultra becomes the default open planning model
Two enterprise agent platforms adopted NVIDIA's Nemotron 3 Ultra on the same day, which makes it a trend, not a coincidence. Glean now orchestrates it among 30+ models, claiming the open model hits 91% of frontier-LLM completeness at open-model cost; its own Waldo search model is post-trained on Nemotron 3 Nano for 50% lower latency. AibleClaw uses Nemotron 3 Ultra for planning in governed, long-running agents, paired with NVIDIA Cloud Functions for a claimed up-to-200x lower agent TCO. The signal for buyers: the enterprise agent layer is going aggressively model-agnostic, swapping frontier-priced calls for open models on the steps where "good enough" is genuinely good enough. Cost, not capability, is the new battleground.
Source: Glean · AibleClaw
Tool of the Day
Walrus Memory
What it’s for: giving your agents cross-session, cross-provider memory without locking it inside any one vendor.
Verdict: Deploy.
Read today’s lead and the Dreaming V3 item back to back and you’ll see the gap Walrus Memory is aiming at. Google can wall off a CLI; OpenAI can rewrite your memories inside its own walls. Both are someone else’s platform. Walrus Memory is a portable memory layer — encrypted, permissioned, and built to be carried across Claude, ChatGPT, and Gemini rather than trapped in one.
It comes from the ex-Meta team behind Sui, ships Python and TypeScript SDKs, and supports proprietary models alongside the OpenClaw and NemoClaw open frameworks. Memory is permissioned, so you control which agents read which slices of context — useful the moment you have more than one agent touching the same user. The free tier is live now, which means this is a real “wire it in this week” tool, not a waitlist.
The honest caveat: a neutral memory layer only pays off if you actually run agents across providers or expect to. If your whole stack is one vendor today, the native memory primitive may be enough. But if avoiding lock-in is a design goal — and after today’s lead, it should at least be a question — this is the layer that’s yours instead of theirs. Start with one agent, point it at the free tier, and see whether portable memory changes how you architect the rest.
Source: PR Newswire
Worth a Click
🔀 Perplexity’s permission-gated local-cloud orchestrator. Perplexity detailed an “air-traffic controller for AI tasks” that auto-routes work between a user’s device and cloud frontier models — and asks permission before sending sensitive work up to the cloud. That permission gate speaks straight to the data-governance anxiety around agentic AI, where the scary part isn't the model, it's what leaves the building. Builds on the June 2 Computex reveal; slated to reach Perplexity Computer in July 2026. marktechpost
💰 Lassie raises $35M for the boring-workflow thesis. a16z led a $35M Series A (about $47M total) for Lassie, whose agents log into insurance portals, pull reimbursements, reconcile them, and update systems of record. Already in 700+ practices across 49 states, claiming 250k+ labor-hours saved a year. The takeaway for builders: deep, unglamorous, vertical workflow agents still pull top-tier checks — you don't need a frontier model, you need a process nobody wants to do by hand. techstartups
The Bottom Line
Today's stat: under 2,000 lines — how little generated code it takes to burn a week of Antigravity CLI quota, per early users. The spicy take to go with it: the most valuable thing in your agent stack this year isn't the model, it's the memory and the tooling you can pick up and walk out the door with. Two of today's three lead stories were vendors quietly tightening their grip. The third was a startup handing it back. Build accordingly.
See you tomorrow.
The Agent Stack · The 5-minute daily brief on B2B AI automation. Unsubscribe · View in browser
| 
