Seroter's Daily Reading — #769 (April 22, 2026) — Seroter's Daily Reading

Listen: https://blossom.nostr.xyz/f0c65d574d3bc9b4f69909e42c6cea6728358d3a56431cd2fac40d781e7098a1.mpga

Seroter's Daily Reading, Episode 769, April 22, 2026. It's day one of Google Cloud Next in Vegas, and as you'd expect, there's a massive amount of news to cover. Richard planted the flag on some key themes from the main keynote: we are in the agentic era, the cloud needs to drive itself, and Google is betting hard on infrastructure that lets agents troubleshoot, remediate, and act across their entire stack. Let me walk you through the highlights.

Let's start with the big picture piece: "Welcome to Google Cloud Next '26." This post frames everything around what Google is calling "the Agentic Data Cloud." The core idea is that data architecture needs to shift from reactive archives built for human scale to proactive systems of action built for trusted agents running at agent scale. They announced cross-cloud lakehouse support, standardized on Apache Iceberg, so you can leave your data in AWS or Azure while querying it through Google. There's a Lightning Engine for Apache Spark that's up to 4.5 times faster than open source. And Knowledge Catalog now acts as a unified context graph, with Gemini autonomously tagging and mapping relationships across your entire enterprise. On the security side, they've introduced Google Cloud Fraud Defense, the evolution of reCAPTCHA, which now handles bots, humans, and agents. And they've deepened the Wiz partnership with specialized agents: the Red Agent hunts for exploitable risks, the Blue Agent investigates threats, and the Green Agent handles root cause analysis and remediation guidance. Deloitte is reporting sixty percent efficiency gains in security operations. University of California Riverside cut incident response from twenty minutes to under two minutes. Lloyds Banking Group completed a full SOC transformation in eight months.

The second piece, "The New Gemini Enterprise" is about the platform for building, orchestrating, and governing agents. This is where it gets serious. Gemini Enterprise now has an Agent Platform with a graph-based framework using the Agent Development Kit. They're introducing Agent Identity, giving every agent a unique cryptographic ID so you can trace and audit everything. Agent Gateway acts as air traffic control between agents and data. And Model Armor protects against prompt injection, tool poisoning, and sensitive data leakage. They've also launched Agent Simulation so you can stress-test agents against real-world scenarios before shipping. And Memory Bank and Memory Profiles let agents maintain context across sessions, which is essential for long-running workflows.

The third piece goes deeper on "Gemini Enterprise for the Agentic Task Force." Here we're talking about long-running agents that can execute complex multi-step workflows over hours or days, managed through a unified Inbox command center. There's a new no-code Agent Designer where anyone can build agents with natural language. Projects creates shared workspaces for teams and agents, and Canvas provides an interactive editor for co-creation across Docs and Slides. They've also added Microsoft 365 interoperability, so you can export to common Office formats. The governance model is baked in from day one.

The fourth piece is "Announcing Spanner Omni." This is interesting. Google is taking Spanner, their globally consistent distributed database, and letting you run it outside of Google Cloud. You can deploy it in your own data center, across clouds, or even on a laptop. It runs on VMs, Linux containers, or Kubernetes clusters, and they're targeting air-gapped and regulated environments where data sovereignty is a concern. For SaaS vendors, this means a consistent technology stack wherever their customers operate.

The fifth piece covers "What's New with Databases." The headline here is the AlloyDB Lakehouse federation preview, which lets you query live data from Iceberg and BigQuery directly from PostgreSQL without data movement. Datastream replication to Iceberg tables is now GA. Spanner's Columnar Engine is GA too, accelerating analytical queries up to two hundred times on live operational data. And Database Center now monitors BigQuery alongside your operational databases, with Gemini-powered fleet analytics surfacing optimization opportunities proactively. CME Group is quoted saying their developers and AI agents can validate, optimize, and innovate in real time.

The sixth piece is "What's New in Cloud Run." Cloud Run's remote MCP server is now generally available, making it easy for developers or agents to deploy code. They're adding billing caps so you can define maximum spend per month. Full-stack apps built in Google AI Studio now deploy to Cloud Run with a single click. And there's a new Cloud Run instances primitive for creating individual instances, ideal for hosting long-running background agents. Replit is mentioned as powering over a million live projects on Cloud Run.

The seventh piece covers "What's New in GKE." Kubernetes is becoming the operating system for the agentic era. Two-thirds of organizations now rely on Kubernetes to power generative AI apps and agents. GKE Agent Sandbox provides secure, low-latency infrastructure for running untrusted code at scale: three hundred sandboxes per second at sub-second latency, with thirty percent better price-performance on Axion compared to other hyperscalers. Lovable is running AI-generated applications in GKE Agent Sandboxes, creating two hundred thousand new projects daily. GKE hypercluster lets a single Kubernetes control plane manage a million chips across two hundred fifty-six thousand nodes spanning multiple regions, backed by Titanium Intelligence Enclave for hardware-attested pod-level isolation. On inference, they've cut time-to-first-token latency by seventy percent with predictive latency boost. Node and pod startup is four times faster. Pod startup times are slashed by up to eighty percent. And intent-based autoscaling now responds in five seconds instead of twenty-five.

The eighth piece is "What's Next in Google AI Infrastructure." This covers the compute layer: new Axion N4A CPU instances delivering better price-performance for agent runtimes. The Virgo Network is their new data center scale-out fabric with four times the bandwidth of previous generations. Storage-wise, Managed Lustre delivers ten terabytes per second of bandwidth, a ten-times improvement over last year. Rapid Buckets on Cloud Storage provides sub-millisecond latency and twenty million operations per second. GKE Inference Gateway uses machine learning for real-time capacity-aware routing.

The ninth piece is a deep dive on "Inside the Eighth-Generation TPU." They introduced two distinct systems: TPU 8t for pre-training and embedding-heavy workloads, and TPU 8i for real-time serving. TPU 8t uses native FP4 floating point to double matrix multiply throughput while maintaining accuracy. SparseCore handles the irregular memory access patterns of embedding lookups. Both systems integrate Arm-based Axion CPU headers to handle data preprocessing and orchestration so TPUs stay fed. For world models that agents use for simulation and reasoning, TPU 8t is optimized to enable millions of agents to practice and refine in diverse simulated environments.

The tenth piece is about "Level Up Your Agents: Announcing Google's Official Skills Repository." This is the work of Richard's team. Skills are compact, agent-first documentation written in Markdown that agents load only as-needed to avoid context bloat. Google launched github.com/google/skills with thirteen skills covering AlloyDB, BigQuery, Cloud Run, Cloud SQL, Firebase, Gemini API, and GKE, plus Well-Architected pillars for Security, Reliability, and Cost Optimization, and recipe skills for onboarding, authentication, and network observability.

The eleventh piece is a change of pace: "15 Principles for Managing Up" from Wes Kao. Managing up has a bad reputation, but the argument here is that anyone can be a leader at any stage of their career, and managing up is a core part of that. Key principles include focusing on the punchline instead of burying the insight, showing your thought process to invite productive pushback, flagging potential issues early with your suggested solution, and embracing that over-communication might actually be the right amount. The piece also reframes micromanagement: often, if you feel micromanaged, it might be that you're not communicating proactively enough. Trust is built by consistently following through and keeping your manager in the loop on the good, the bad, and the ugly.

The twelfth piece is "Eclipse Foundation Offers Enterprise-Grade Open Source Alternative to Microsoft's VS Code Marketplace." Open VSX is the vendor-neutral extension registry for VS Code-compatible tools. The Eclipse Foundation is now offering a managed registry with a ninety-nine point nine five uptime SLA, service credits, and defined support tiers. Amazon, Google, and Cursor are initial customers. The registry serves over three hundred million downloads per month with peak daily traffic exceeding two hundred million requests. The key insight is that AI-driven development has changed the economics of developer infrastructure. Where registries used to serve human developers, AI agents now create machine-scale traffic that demands production-grade reliability.

The thirteenth piece is "Introducing the Builders Hub from the Google Developer Program." Builders Hub is a new centralized entry point that aggregates developer resources across Google's many consoles and documentation sites. It provides a unified workbench for projects and personalized suggestions for community engagement and learning. The friction of jumping between surfaces has been a real pain point, and this consolidates that.

The fourteenth piece is a wild one from TechCrunch: "SpaceX is Working with Cursor and Has an Option to Buy the Startup for Sixty Billion." SpaceX struck a deal with Cursor to develop next-generation coding and knowledge work AI. The deal includes an option to acquire Cursor for sixty billion later this year. Cursor was valued at two point five billion in January of last year, then nine billion by last May, and twenty-nine point three billion in its Series D in November. That's a steep climb. Interestingly, two of Cursor's senior engineering leaders already left to join xAI last month. SpaceX's partnership with Cursor combines their product and distribution with the Colossus supercomputer, which claims the compute equivalent of a million Nvidia H100 chips. Neither Cursor nor xAI has proprietary models that match Anthropic or OpenAI, and this partnership may be designed to eventually escape that dependency.

The fifteenth piece is "A Guide to Five Agent Payment Protocols." This is a dense technical piece on how agents will pay for things autonomously. Google and partners created AP2, which uses cryptographically signed mandates to authorize agents to act on your behalf. OpenAI and Stripe created the Agentic Commerce Protocol, which launched in ChatGPT's Instant Checkout. Coinbase created x402, which leverages the HTTP 402 Payment Required status and accepts only cryptocurrency. Stripe and Tempo created Machine Payments Protocol, which also uses HTTP 402 but accepts credit cards and debit cards in addition to cryptocurrency. And the Agent Transfer Protocol introduces agent-specific intent verbs like QUERY, BOOK, SCHEDULE, and PURCHASE. Each protocol has different tradeoffs around autonomy, security, payment method support, and whether they're designed for real-time or user-mediated transactions.

The sixteenth piece is from HR Dive: "Employers Say They Struggle to Find Workers with the Right AI Skillset." Despite AI changing how companies operate, fifty-three percent of employers report difficulty finding graduates with the right AI skills. Seventy-eight percent of higher education leaders believe they're meeting employer expectations, but only twenty-eight percent of employers agree. Only fourteen percent of graduates have achieved high proficiency in applying AI tools professionally. Meanwhile, eighty-three percent of workers believe AI can perform most entry-level jobs. The gap between what universities are teaching and what the market needs is significant, and Tom ap Simon from Pearson put it well: basic AI literacy is no longer sufficient. Schools that lead in AI readiness today will shape the future of workforce readiness tomorrow.

Across all these pieces, a few tensions emerge. First, the infrastructure layer is racing ahead of the governance and security frameworks needed to operate safely at scale. Google is moving fast on agents, but the emphasis on Agent Identity, Agent Gateway, and Model Armor shows they know sprawl and risk are real concerns. Second, the skills gap is real. Universities aren't keeping pace, and enterprises are feeling it. This episode is heavy on Google Cloud Next, but that workforce readiness problem is going to be a limiting factor for every organization trying to move faster into the agentic era.

That's Episode 769. Thanks for listening, and I'll see you tomorrow for day two.