Seroter's Daily Reading — #787 (May 19, 2026) — Seroter's Daily Reading

Listen: https://blossom.nostr.xyz/9adbf1bb083f5f937d45733030d82d71ad606715d7be31f0b0c7a1295b5f2251.mpga

Seroter's Daily Reading, episode 787, May 19, 2026.

Whew. The workday isn't done yet, but I'm getting a quick breather from Google I/O going on this week. The keynotes were terrific, and I just finished a breakout with two superstar presenters. Below you'll find a few I/O items I read, and then non-Google stuff.

Let's start with the big picture from Google I/O. The central theme this year is clearly the agentic era, and they announced some genuinely new ground with I/O 2026: Welcome to the agentic Gemini era. The most interesting consumer-facing announcement is information agents in Search — personalized AI agents that run in the background 24/7, tracking things you care about, alerting you at exactly the right moment. If you're a Google AI Pro or Ultra subscriber, these start rolling out this summer. But Search is going further than simple alerts. With generative UI powered by Gemini 3.5 Flash and something they're calling Antigravity, Search can now build custom dashboards and trackers for longer-running tasks. Think of these as mini apps generated on the fly for your specific questions. You can build them right in Search, also coming to Pro and Ultra subscribers first.

The other headline is Gemini 3.5: frontier intelligence with action, now available. It's the speed-optimized model in the family, and the pitch is that it handles long-horizon agentic tasks faster and cheaper than other frontier models — often less than half the cost. The pairing with Antigravity turns it into a collaborative subagent engine for demanding workflows. 3.5 Flash plans, builds, and iterates across multi-step tasks, and under supervision can reliably execute code and workflows while staying at frontier performance levels.

Then there's Introducing Gemini Omni, which was the most impressive thing I saw from the keynotes. Building on the Nano Banana image model from last year, Omni takes multimodal input — images, audio, video, text — and generates high-quality video grounded in Gemini's real-world knowledge. You can edit video through conversation, with every instruction building on the last. Characters stay consistent, physics hold up, and the scene remembers context. The first model, Gemini Omni Flash, is rolling out now in the Gemini app, Google Flow, and YouTube Shorts. Support for image and audio output is coming later.

Shifting to the developer-focused announcements from Building the agentic future: Developer highlights from I/O 2026 — Google shipped a major refresh to Antigravity, and it's worth looking at what they're building here. Antigravity is their agent-first development platform, and this week brought a full desktop application, a CLI, an SDK, and enterprise integration. The desktop app acts as a central hub for orchestrating multiple agents in parallel, with dynamic subagents, scheduled background tasks, and integrations across AI Studio, Android, and Firebase. The CLI is for terminal-preferring developers who want to create agents without a GUI. The SDK gives programmatic access to the same harness powering Google's own agents. And enterprise customers can now connect Antigravity directly to Google Cloud projects. There's also a new $100 per month AI Ultra plan with five times higher Antigravity usage limits, and a limited-time offer of $100 in bonus credits that expires May 25th.

Also new in the developer stack: Managed Agents in the Gemini API. One API call spins up an agent that reasons, uses tools, and executes code in an isolated Linux environment. Each interaction creates a persistent environment you can resume in follow-up calls, with all files and state intact. You can define custom agent behaviors with markdown files and get started with templates in Google AI Studio. And on the AI Studio side, there's a new mobile app for capturing ideas on the go, Workspace API integration for agents to call Google tools natively, one-click export to Antigravity for local development, and native Android app building from prompts. Google also launched the Build with Gemini XPRIZE Hackathon, with two million dollars in prizes — the largest ever for a hackathon. Developers can build real applications solving challenges from reducing food waste to advancing scientific research, with finalists pitching live at the Moonshot Gathering in Los Angeles this September.

On the enterprise side, Everything Google Cloud customers need to know coming out of Google I/O makes clear that many of these announcements aren't just for consumers — Google Cloud customers get access too. Gemini Spark, the 24/7 personal agent, comes to Gemini Enterprise. It works in the background across Workspace, custom connectors, and the open web, executing multi-step workflows on your behalf with explicit approval required for high-risk actions. It learns your preferences over time and connects to existing Gemini Enterprise connectors including Microsoft Sharepoint, OneDrive, and ServiceNow. The security model runs every task in a fresh, strictly isolated, ephemeral VM, with traffic routing through an Agent Gateway that enforces Data Loss Prevention policies. The blog has quotes from Accenture, AirAsia Next, Deloitte, Monks, PwC, and WPP — all describing how Antigravity has shifted their teams from manual coding to high-level orchestration, with some claiming more than half their production-ready code is now generated through agentic workflows.

Now for something completely different. There's a substantial piece from O'Reilly on Agent Skills Work but the Research Shows Most Teams Are Building Them Wrong and what the research actually shows about whether teams are building them wrong. Four recent papers take the first systematic look at skills in practice, and the findings are worth sitting with. Curated skills — written by humans from real execution experience — raised agent task success rates by 16.2% on average across 84 tasks. Model-written skills showed no consistent benefit. As skill libraries grow, flat retrieval breaks down. Similar-sounding skills start colliding, and the paper calls the failure mode routing collapse, where the agent consistently invokes the wrong skill because the embeddings are indistinguishable. The fix is organizing skills into a capability hierarchy — top-level domains like code, data, docs, with more specific skills as branches and leaves. The agent navigates domain to branch to leaf instead of scanning everything.

A large-scale security study of over 31,000 community skills found that more than one in four contain exploitable vulnerabilities — prompt injection, data exfiltration, privilege escalation. More than a quarter. This changes what importing a skill from a public repository should look like. It's like doing an npm install from an unknown author. You wouldn't do that without checking what the package does. The research recommends three things: write skills from real execution, not from scratch; treat the skill description as routing logic, not a label — it determines whether the skill fires at all; and plan for the full lifecycle, because a skill that compensated for a model capability gap six months ago may now be actively overriding better native behavior.

On the deployment side, Will Denniss has a walkthrough for Deploying to Agent Platform with ADK for deploying an agent built with the ADK to Google Agent Platform. The script configures your project, staging bucket, and location, then calls the deployment API with your agent configuration. One note — for production Google Cloud calls, set GOOGLE_GENAI_USE_VERTEXAI to TRUE and don't pass in GOOGLE_API_KEY. The agent's service account identity handles auth automatically. He also shows how to interact with the deployed agent via HTTP, using gcloud to get an access token and then streaming queries through the session API.

A quick note on a DZone piece about multi-cloud lessons — the article was blocked by a security service, so I can't actually tell you what it says. Sometimes the research hits a wall.

Finally, a thought-provoking piece from Sean Goedecke on The just-say-no engineer was a ZIRP phenomenon and why this archetype is struggling. His argument: it's not about AI at all. It's about the end of ZIRP — the zero interest rate policy era from 2008 to 2022. During that period, investors were throwing borrowed money at anything, tech companies were constantly hiring, and having a very senior engineer whose only job was to say no was actually valuable. They kept systems from becoming unmanageable when half the company was proposing changes and being told no. Having a reputation for high technical standards was a positive for hiring. Now that interest rates have risen, companies have to actually make money. They're not doing random crap anymore. They're chasing features that can generate revenue. That environment is actively inimical to the just-say-no engineer. Management support is gone. They get bad reviews for the exact same behavior that was rewarded pre-2022. AI adds insult to injury — they're watching other engineers merge AI-generated PRs that would have been blocked, and the AI tooling mostly works. The code isn't quite as clean, but it's good enough, particularly in a world where companies are trying lots of new things and abandoning the ones that fail. The just-say-no engineer isn't going extinct — there are domains like pure engineering work where they remain essential. But their scope has contracted, and that has nothing to do with language models.

That's episode 787. Big theme across the week: the agentic era is here, and the tooling is maturing fast — from Antigravity to Managed Agents to Gemini Omni. On the skills research front, the takeaway is that curated beats generated, hierarchy beats flat retrieval, and you should review imports before loading them. And on the career dynamics side, the ZIRP era is over for everyone, including the gatekeepers. See you next time.

Sources: