
Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Google’s big developer pitch is “managed agents” — via the Gemini API, developers now get an agent plus a remote Linux sandbox in one call, so Google handles the “infrastructure nightmare” while teams like Stitch connect to GitHub repos, analyze codebases, and generate design artifacts at scale.
AI Studio is being positioned as prompt-to-app, then prompt-to-deploy — Paige showed a Hacker News radio app built around a managed agent, deployed to Cloud Run in a few clicks, with Google saying new users can ship to a live URL with no credit card required.
Google is leaning hard into agents as the new programming surface — Logan’s demo defined skills and tools in markdown for an “AI talk radio” agent that researched Hacker News, wrote a script, generated voices, music, cover art, and mixed an MP3, prompting his line that “the hottest new programming language is Markdown.”
The anti-gravity stack now spans desktop, CLI, SDK, and multi-agent orchestration — Google announced anti-gravity 2.0, the anti-gravity CLI, one-click export from AI Studio, dynamic sub-agents, and scheduled tasks, all aimed at making agents a day-to-day development workflow instead of a novelty demo.
Wes Roth’s real reaction: the tooling sounds useful, but the presentation lost steam — he cut away early, saying he was “very, very bored,” but still called the managed-agent Linux VM setup and the CLI genuinely interesting enough to test against tools like Codex, Cloud Code, and even Grok Build.
The practical takeaway for Wes was less “wow demo,” more “I want hands-on time” — he said the most exciting part was not watching Google walk through it, but trying the features himself, especially the hosted Linux sandboxes, AI Studio Android app generation, and developer integrations like Firebase, Workspace, and GitHub.
The stream opens with Wes fumbling a bit with playback, then Google jumps straight into the “where we were a year ago” framing and rattles off the headline stack: Gemini 3.5, Gemma 4 under Apache 2, and a broader shift from assistants to agents that actually get work done. The centerpiece is “anti-gravity,” Google’s agentic development platform, which they position as the connective tissue across models, tools, platforms, and infrastructure.
The first concrete launch is managed agents in the Gemini API. Logan frames the pitch very cleanly: building the agent is only half the battle, because in production you also need secure, isolated execution — so Google now pairs each managed agent with a remote Linux environment it hosts for you. The example was Stitch, Google Labs’ vibe-design tool, where an agent connects to GitHub, inspects a codebase, and outputs a design file that helps generate on-brand designs.
Logan’s demo is an agent template inside AI Studio that turns a topic into a full talk-radio episode. This one pulled top Hacker News stories, wrote a script, generated multiple TTS voices, made music with Lyria, created cover art with Nano Banana, mixed everything into an MP3, and dropped the finished file into its sandbox. His punchline was the memorable bit: he didn’t write orchestration logic at all — just markdown files defining skills and tools — so “the hottest new programming language is Markdown.”
Paige takes that same radio-show concept and wraps it in an actual app, showing the shift from toy demo to something shippable. The managed agent creates the MP3 plus metadata for speaker navigation, then AI Studio pushes the app to Cloud Run in a few clicks, with a no-credit-card path for new users. She also runs through broader integrations: Firebase, Firestore, Google Workspace, web search, and image generation, all sold as ways to stay “in the flow.”
The next push is AI Studio as a full app-building front end, not just a prompt playground. Paige shows Android app generation in Kotlin, previews it in an emulator directly inside AI Studio, then talks through publishing to Google Play test tracks and a coming mobile app for AI Studio itself. The theme is obvious: no local setup, no SDK wrangling, just prompt your way from idea to app.
Anel shifts the presentation toward anti-gravity 2.0, now described as an “unabashedly agent first” desktop app. The new features are all about orchestration: multiple agents across multiple projects, dynamic sub-agents like QA or data-science helpers, and scheduled tasks so agents can summarize PRs or monitor cloud health on autopilot. Google also announces a one-click export from AI Studio into anti-gravity and an anti-gravity CLI for developers who live in the terminal.
Kevin starts a demo about fine-tuning Gemma 4 with LoRA so a model returns only bash commands — useful for self-healing CI pipelines instead of verbose conversational output. But this is where Wes cuts in with his own reaction: “I am very, very bored,” and he ends the watch-along early, telling viewers to continue on Google’s livestream if they want the full presentation. It’s a very Wes moment — blunt, a little apologetic, and powered partly by a “very large poke bowl.”
After stopping the stream, Wes still recaps what mattered to him: Gemini 3.5 Flash and Pro, anti-gravity 2.0, the CLI, SDK, scheduled tasks, and especially the isolated Linux sandbox environments behind managed agents. He says he wants to benchmark the CLI against Codex and Cloud Code, notes Grok Build as another comparison point, and basically lands on this: the announcements are probably valuable for developers deep in the stack, but the real story will be what happens once people actually get their hands on the tools.
Share
Keep Reading
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.

Playbook
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.

Playbook
Learn how tasteful prompting helps you move beyond generic AI output by shaping context, style, and judgment from the start.

Playbook
OpenAI shipped /goal for the Codex CLI. It turns a prompt into a persisted, self-continuing contract.