In partnership with
Hello from The AI Night,
Today in AI:
Google Launches Gemini Desktop App for Mac
OpenAI Agents SDK Gets Sandboxes, Memory, MCP Support
Google Launches Fabula, an AI Writing Tool for Authors
Here's the deal: Google released a Gemini desktop app for Mac. Users can open it from any screen using Option + Space and share the active window so Gemini can answer questions based on what is currently on screen.
The Breakdown:
Access point; Option + Space pulls up Gemini from any app without switching windows.
Window sharing; users can share a specific window, letting Gemini see documents, code or data in context.
Use cases cited by Google: working with documents, code and data.
Platform; Mac desktop, adding to Gemini's existing web and mobile surfaces.
The announcement does not specify supported macOS versions, pricing or whether the app is gated behind a specific Gemini plan.
The bigger picture: Window sharing is the real story. Gemini can now see what you see without a screenshot, which is how Google quietly catches up to Anthropic's computer use and OpenAI's agent mode. The Mac app is the wrapper. Live screen context is the actual capability shift.
Here's the deal: OpenAI shipped an update to the Agents SDK adding a new harness for the agent loop and native sandbox execution, so agents can run code, edit files, and use tools in controlled environments.
The Breakdown:
The harness bundles configurable memory, Codex like filesystem tools, MCP, skills, AGENTS.md, a shell tool and an apply patch edit tool.
Native sandbox support ships with Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop and Vercel or developers can bring their own.
A Manifest abstraction defines the agent's workspace and pulls data from AWS S3, Google Cloud Storage, Azure Blob Storage or Cloudflare R2.
Harness and compute are separated, enabling snapshotting so runs survive lost containers, plus parallel subagents across isolated sandboxes.
Generally available via the API at standard token pricing. Python first with TypeScript, code mode and subagents planned.
The bigger picture: Every team building production agents has quietly rebuilt the same sandbox and memory stack. OpenAI just made that work disposable. The real shift is snapshotting, which means agents can finally survive crashes and run for hours without losing state. That is the wall every serious agent team hits first.
Here's the deal: Google Research announced Fabula, an interactive AI writing tool built to help authors structure and refine stories. The tool was co-designed with 42 expert writers and is being showcased at a live demo scheduled for 10:30AM at the Google booth.
The Breakdown:
Fabula is positioned as an interactive writing tool focused on structuring and refining stories per Google Research
Co-designed with 42 expert writers, which Google frames as core to the tool's development process
The demo centers on "convergent iteration," which Google describes as a method that supports creativity during writing
The announcement was made via Google Research's official X account alongside the live demo event
The source does not disclose model details, availability, release timeline, pricing or the hosting venue
The bigger picture: Most AI writing tools help you generate faster. Fabula is built to help you write better, which is a harder problem that usually means worse adoption numbers. Co-designing with 42 writers signals Google is optimizing for craft not speed. The real test is whether serious authors will actually trust it.
LLM traffic converts 3× better than Google search
58% of buyers now start their research in ChatGPT or Gemini, not Google. Most startups aren't showing up there yet.
The ones that are get cited by the AI tools their buyers, investors, and future hires already use. And they convert at 3×.
Download the free AEO Playbook for Startups from HubSpot and get the exact steps to start showing up. Five minutes to read.
What else you need to know:
Google launched Skills in Chrome, letting users save frequent Gemini prompts as one-click workflows that run on the current page and selected tabs, with a prebuilt library included.
Windsurf launched 2.0 with an Agent Command Center that tracks local and cloud agents in a Kanban view, and bundled Devin, its autonomous cloud engineer, into every plan.
Mathematician Jordan Lichtman says GPT-5.4 Pro solved Erdős #1196, a 1966 primitive-set conjecture, by keeping analytic von Mangoldt weights rather than switching to probability as prior work did.
NVIDIA Research introduced Lyra 2.0, a framework that generates persistent, explorable 3D scenes from a single image and exports them as Gaussian Splats or meshes to Isaac Sim for embodied AI.
Google added subagents to Gemini CLI, letting the main session delegate tasks to specialized agents with isolated context windows, custom tools, and system instructions, executable in parallel.
That’s it for today’s edition of The AI Night.
Our goal is to cut through the noise, surface what actually changed, and explain why it matters.
3 ways to support us:
Forward this to your AI-curious friend → https://www.theainight.com
Sponsor The AI Night and reach 500+ AI builders daily → passionfroot.me/theainight
Reply to this email — I read every response