Hey Everyone 👋,
John Lindquist here with the 39th issue of AI Dev Essentials!
I/O week. Google spent two hours showing how deeply it wants agents wired through its models, developer tools, Workspace, Search, and media products. Gemini 3.5 Flash, Antigravity 2.0, Gemini Spark, Gemini Omni, and CodeMender all showed up across the I/O announcement stream. The week wasn't only Google though. Anthropic announced its acquisition of Stainless, OpenAI ran another Codex Thursday, and Claude Code kept polishing how plugins and sessions are tracked.
This issue is the Google I/O Roundup, with the rest of the week's big developer stories underneath. The pattern keeps showing up. Every major lab is racing to put an agentic system between you and your editor, and the practical advantage goes to developers who learn where these systems actually fit inside real codebases.
That is exactly what next Friday's Codex Power User Workshop is built around. One week from today. This first Codex-focused cohort is about turning Codex into a repeatable development system: context packaging, parallel terminal agents, tool profiles, skills, hooks, durable memory, and the Codex SDK.
Full curriculum is at the bottom. Let's get into I/O.
⚡ Codex Power User Workshop — 📅 Fri May 29, 9am PT Reserve your spot →
📚 dev.build Lessons
Codex Plugins
I'm recording lessons on building Codex plugins which should be posted later today.
🚀 Google I/O 2026 Roundup
Model Update Google Ships Gemini 3.5 Flash as the New Agent Default
Google announced Gemini 3.5 Flash at the I/O 2026 keynote on May 19, 2026, positioning Flash as the default engine for long-horizon agentic work rather than just a cheap-and-fast tier. It is rolling out across Search, the Gemini app, and the Gemini API, with Gemini 3.5 Pro slated for next month.
Highlights:
- Surpasses Gemini 3.1 Pro on coding and agentic benchmarks
- Tuned for plan, build, iterate, tool use, code execution loops
- Live now in the Gemini API and the AI Gateway as
google/gemini-3.5-flash - Available across the consumer Gemini app and Search AI Mode
(Google — I/O 2026 Keynote Recap, Gemini 3.5 Flash on the AI Gateway)
💬 3.5 Flash is a mixed bag. It is fast and powerful, but feels like it's focusd on one-shot tasks rather than long coding sessions. It's had a really difficult time self-correcting and you'll see it burn through your entire quota going down the wrong path. I do feel like I just haven't figured out how to use it, so I'm hesitant to call it "bad". There are a ton of scenarios where you want a "quick fix" rather than an overnight "/goal" session. I just need to figure out what those are. I'll keep y'all posted as I work through these scenarios.
New IDE Antigravity 2.0 Launches as a Standalone Desktop App and CLI
Google launched Antigravity 2.0 at I/O 2026 as a full standalone desktop app for Mac, Windows, and Linux, with a new Antigravity CLI that replaces the older Gemini CLI. It is positioned as Google's flagship agent-first development environment built on Gemini 3.5 Flash.
What's New:
- Standalone desktop app with autonomous agents that plan, write, test, and verify
- Antigravity CLI for headless and CI workflows, replacing Gemini CLI
- Weekly Gemini quotas tripled twice in the launch week after early users hit limits
- Tiered offering: a low-friction individual tier and a governed enterprise tier through Gemini Enterprise
(Google — Antigravity 2.0, Antigravity CLI Docs)
💬 I'll be honest, I prefer TUIs and it feels like all of the IDEs eventually just copy Cursor. I have been using the AGY TUI fairly extensively and it's "fine". It's lagging pretty far behind Codex and Claude Code, but I think it will catch up quickly. VGoogle posted that they tripled weekly quotas, reset them, and then tripled again because people were burning through limits in a couple of work sessions. I definitely noticed burning through tokens like crazy, so hopefully this will at least help test it out more.
Ecosystem Gemini Spark Brings an Always-On Personal Agent
Gemini Spark is Google's new personal agent that takes actions on your behalf across Gmail, Docs, Calendar, and Workspace, running in the cloud even after you close your laptop. Third-party MCP server support is slated for this summer. Google says Spark is in trusted testing now, with a beta planned for Google AI Ultra subscribers in the U.S. next week.
Capabilities:
- Cloud-resident agent that keeps running when your laptop is closed
- Crosses Gmail, Docs, Calendar, Drive, and Workspace
- Third-party MCP servers supported later this summer
- Pairs with Daily Brief, a morning digest that scans Gmail, Calendar, and Tasks and suggests next steps
(Google — Gemini Spark, Gemini)
💬 I use Google for everything: Gmail, Docs, Calendar, Drive, etc, etc. So when I set up an agent like Hermes or OpenClaw, most of the work goes into it managing my Google accounts. So this is super exciting (kind of). I just don't really trust Google to pull this off for power users with multiple Google accounts and I doubt you'll have many customization options. But I'm guessing this will be a great solution for people who aren't interested in managing their own agents.
Model Update Gemini Omni Starts With Multimodal Video Through Omni Flash
Google introduced Gemini Omni at I/O 2026 as a model for creating outputs from text, image, audio, and video inputs, starting with video. Gemini Omni Flash is the first rollout across the Gemini app, Flow, YouTube Shorts Remix, and YouTube Create, with developer API access coming in the next few weeks.
Notable:
- Voice-driven editing: describe a change, get a re-rendered cut
- Multimodal inputs: text, image, audio, video
- Pairs with Veo 4 for higher-end generation and Flow for editing
- Developer and enterprise API access is rolling out through the Gemini API and Agent Platform API in the coming weeks
(Google DeepMind, Google — AI Blog)
💬 I uploaded my own avatar and did a couple of test videos. It looks and sounds like me, but we've seen this trick already with Sora. And now video models like SeedDance are doing almost cinematic-quality video generation. So while I fully believe Omni will be an amazing model, the video generation aspect feels far behind the state of the art. And my general disappointment with I/O is that nothing that they launched was best-in-class. Everything felt a few months behind things we're already using.
Security Google Folds CodeMender Into the Agent Ecosystem
Google expanded CodeMender at I/O 2026 by integrating the AI code security agent into Agent Platform. Google says it can identify vulnerabilities, recommend fixes, test them securely, and apply patches with developer approval.
What It Does:
- AI-driven vulnerability triage and patching at scale
- Tests proposed fixes in a secure environment before applying
- Hooks into Google's Agent Platform alongside Gemini Enterprise
(Google — Safety & Security Blog, CSO Online)
💬 If 2025 was the years of PR Review AI tools, 2026 is the year of AI security tools. And, holy crap, this year has been a security nightmare so far. Whatever happens next is anyone's guess.
Tool Managed Agents and AI Studio Round Out the Developer Story
Beyond the headline models, Google shipped Managed Agents in the Gemini API and a set of AI Studio upgrades during the I/O developer keynote on May 19, 2026.
For Developers:
- Managed Agents in the Gemini API for provisioning an Antigravity-powered remote Linux environment where agents can reason, call tools, execute code, manage files, and browse from an isolated sandbox
- Google AI Studio upgrades for Android app generation, Workspace-connected apps, Google Play test-track publishing, Antigravity export, and preview-window annotations
- Android XR Developer Catalyst Program with pre-release access to XREAL's Project Aura, audio glasses, and display glasses
(Google AI Studio, Google — AI Blog, Android XR Catalyst)
💬 I need to get my family playing on AI Studio. It's probably my favorite Google AI tool, but just not something I'd ever personally use.
🚀 Major Announcements
Acquisition Anthropic Acquires Stainless
Anthropic announced its Stainless acquisition on May 18, 2026. Anthropic did not disclose deal terms, though prior reporting put acquisition talks above $300M. Stainless builds SDK, CLI, and MCP tooling used by hundreds of companies, including Anthropic itself.
What It Means:
- Every major Anthropic competitor has used Stainless for SDK generation
- Watch whether hosted Stainless products or competitor SDK workflows change after the acquisition
- Anthropic is signaling that SDK and developer infrastructure is strategic, not a commodity
- Competing labs will likely need to revisit how they generate and maintain their developer SDKs
💬 This is the kind of acquisition that looks boring on the outside and rewires the industry on the inside. SDKs are how every agent framework actually talks to a model. The framing showing up in my feed is blunt: Anthropic just acquired a piece of shared developer infrastructure and is in a position to influence what their competitors ship. Anthropic is being careful about messaging, but the lever is real. Worth watching how OpenAI and Google respond over the next two quarters, and whether they double down on their own SDK generation in-house.
Tool Codex Thursday: Appshots, Remote Computer Use, Goal Mode, and Browser Annotations
OpenAI ran another Codex Thursday on May 21, 2026, shipping a stack of updates focused on cross-device control, screen-aware context, and longer autonomous runs.
Updates:
- Remote computer use: Let Codex use desktop apps after your Mac locks, including remote access through Codex Mobile, with short-lived authorization and relock safeguards
- Appshots: Pull context from your screen straight into the Codex app
- Goal Mode is now available across the Codex app, IDE extension, and CLI for hours-or-days-long autonomous runs
- Advanced in-app browser annotations: Tweak font size, colors, and spacing directly in annotations so Codex gets clearer change requests
Codex CLI 0.133.0 also landed the same day with goals enabled by default, dedicated goal storage, permission profile list APIs and inheritance, managed requirements.toml support, and stronger Windows sandbox integration.
(OpenAI Codex, Codex CLI Changelog)
💬 "/goal" is my new favorite feature and Codex is the only one that really pulls it off so far.
Tool Claude Code Improves Plugin and Usage Visibility
Claude Code's /usage command surfaces session-level token statistics, and recent changelog entries keep adding visibility into plugins and background sessions.
Why It Matters:
- Plugins, Skills, agents, and MCP servers can quietly add context and token cost
- Current-session usage stats help with cost checks, but per-feature attribution still needs a stronger source
- Lands while plan limits and weekly caps are an ongoing conversation across every coding tool
💬 Boris Cherny from Anthropic teased a per-Skill, per-Agent, per-MCP, per-Plugin token breakdown for the next Claude Code release, and the response in my feed was instant. People want this, badly. It is too easy to install five plugins and have no idea which one is the reason you hit your weekly cap on a Wednesday.
🛠️ Developer Tooling Updates
Open Source Roughdraft Brings Comments and Suggested Changes to Markdown Plans
Roughdraft surfaced this week as an open source, local app for collaborating with agents on markdown plan docs. It adds inline comments and suggested changes to plain markdown while keeping the workflow local.
💬 Just one of those projects that caught my eye. Collaborating on Markdown files with humans and agents seems like a good next step, but how is this different from Notion? This space is ripe for innovation if you're looking to build tools for agentic workflows that will have a meaningful impact.
Tool Remotion Publishes AI Coding Agent Guidance and Skills
Remotion now documents AI coding agent workflows for Claude Code, Codex, and OpenCode, with Remotion skills that teach agents project conventions. Combined with HTML-in-Canvas, the project is leaning into agent-driven motion graphics without forcing a separate chat UI.
💬 Remotion is one of the cleanest "let an agent generate a real artifact" stories out there. If Claude can drive the timeline and components reliably, this becomes a serious option for programmatic video. I've done some cool stuff with it a few months ago, but I definitely need to dive back in after the html canvas upgrade.
✨ Workshop Spotlight
Codex Power User Workshop
One week away, first cohort
A full-day, hands-on session that turns Codex from "another CLI" into a repeatable development workflow you can trust on real projects. Leave with a system you can keep running after the session, not just a notes doc full of tips.
Friday, May 29 @ 9:00 AM to 2:00 PM Pacific
- Context Packaging: Bundle files, logs, screenshots, and traces so Codex understands the real problem
- Terminal Workflows: Run parallel research, planner, builder, and validator sessions without losing the thread
- Agent Profiles: Compare broad defaults vs. narrow prompts, focused skills, and limited tool access
- Tools & Sandboxing: Separate read-only from mutation tools and map agents to repeatable tool profiles
- Skills, Hooks & Memory: Build reusable expertise, enforce guardrails, and capture durable decisions
- Codex SDK: Embed Codex into CI, dashboards, and internal tools when terminal sessions aren't enough
Did you enjoy this issue? Share it with a friend.
© 2026 John Lindquist • egghead.io