Seroter's Daily Reading — #773 (April 29, 2026) — Seroter's Daily Reading

Listen: https://blossom.nostr.xyz/cb1fc5f68cc4b42314cc7d450bbd92eccca1172b7d1807831ce50ef1f1c86ef4.mpga

Seroter's Daily Reading, Episode 773. April 29, 2026.

We had quite the earnings call at Google today, and I'm proud of the work Google Cloud has done. But I'm on a bit of a recharge week, so let's get into the reading list.

Leading this week with Google Cloud Next. The blog post recapping the event is titled "260 things we announced at Google Cloud Next '26". That's a lot of things. Looking at this recap, it's a mix of product news, customer stories, and partner updates. The central theme was what Google is calling the Agentic Data Cloud, and they announced a flood of capabilities around Knowledge Catalog, BigQuery enhancements, cross-cloud lakehouse features, and database improvements. There are new agents for data engineering, data science, and database management. They launched new GPU options for Cloud Run, Firebase SQL Connect, and a bunch of GKE updates. The customer stories section shows over 1,300 production AI use cases now. Capcom building gameplay agents, Citi Wealth launching an AI-powered wealth advisor, Deutsche Telekom deploying autonomous network operations. HCA Healthcare handling 80,000 patient handoffs a day with agentic systems. This is the scale we're talking about.

But Keith Class over at The CTO Advisor cut through all 260 announcements and identified the real strategic pattern. His piece is titled "The Industry Is Watching GPU Prices. Google Just Moved the Fight to the Judgment Layer." Keith argues that Google is systematically remixing existing platform capabilities into what he calls the borrowed judgment layers, while the market argues about compute prices. He draws a parallel to the early cloud era. Back then, AWS was happy to sell you bare metal infrastructure, but the real value was the IAM, the managed databases, the governance primitives, the operational logic that became embedded in how organizations ran their workloads. The compute was commodity. The operational model built on top of it was not. Keith says the AI infrastructure market is replaying that exact pattern. The neoclouds are the new colocation providers with real price advantages on compute. The hyperscalers are building the judgment layers above it. And Google's Knowledge Catalog is the key example. Keith describes how he spent ten months building on GCP, struggling with the grounding problem in his own AI system, and the entire time the solution was sitting there under his feet. Dataplex for metadata and lineage, Data Catalog for discovery, BigQuery for semantic relationships, Gemini for intelligence. He didn't see it because those pieces presented themselves as data governance plumbing, not as an AI grounding architecture. Google remixed them into Knowledge Catalog and in doing so made a previously hidden architecture visible. That's the play. Not a new product. A new framing of capabilities that were already there. The cross-cloud lakehouse is the judo move: your data can stay in AWS or Azure, Google brings the judgment layer to your data. The switching costs aren't in storage. They're in the authority model and the governance infrastructure your agents depend on.

Speaking of agents, Google also released the Agents CLI in Agent Platform. This is a unified programmatic backbone for the agent development lifecycle on Google Cloud, designed specifically for AI coding agents like Gemini CLI, Claude Code, and Cursor. One command to inject bundled skills directly into your coding environment. This is what using agents to create agents looks like.

But not everyone's buying what the vendors are selling. A16z published a piece titled "Workday's Last Workday?" This is a sharp analysis of Workday's competitive position. Workday is arguably the most important and least loved product in enterprise software. More than 10,000 organizations run it, approaching $10 billion in annual revenue, and public markets value it at roughly $30 billion. None of that success is because the product is beloved. HR admins spend their days on data entry, workarounds, and hand-holding. When customers renew at close to 100% every year, it's usually read as a sign the product is delightful. In Workday's case, it's a sign of something else: leaving is close to impossible. The article walks through Workday's moat in detail. Deep technical and human wiring. Workday sits at the center of hundreds of integrations, none of which migrate cleanly. Thousands of hours of muscle memory inside each instance. A proprietary configuration layer built in Workday Studio that takes 6 to 18 months to implement. More than 10,500 certified consultants. Multi-year contracts. The combination is why Workday posts some of the best gross revenue retention in enterprise software. But the article argues this moment is different. Three things have changed. Enterprise IT is finally revisiting core systems because AI-readiness reviews are turning legacy architectures into liabilities. The tools to rebuild HCM now exist, including AI-native SAP migrations at Fortune 500 scale. And Workday structurally can't close the gap from inside. Bolting AI on top of their twenty-year-old forms-and-approvals engine doesn't change what's underneath. The piece concludes that HCM is the last large enterprise software category without a serious AI-native challenger, and that's about to change.

Meanwhile, at the other end of the enterprise software spectrum, Bed Bath & Beyond's CEO Marcus Lemonis said on their earnings call that AI will lead to a significant reduction in headcount. He's being brutally clear and honest about it. The company is becoming an organization that puts its payroll in the field, that puts its payroll generating revenue, and does not put its payroll in corporate offices with big leases and lots of warehouses. Bloated back offices are going away. Areas impacted include supply chain, IT, accounting, marketing, and merchandising. In some cases those jobs will be redeployed into customer service and store staff.

On the engineering leadership front, a piece from the manager.dev newsletter looks at managing a team that didn't choose you. The author took over an existing team after a career break, and was told in the first 30 days to only listen and learn. That didn't work out. Three days before they started, the team was moved to a new group, thrown into a completely new project, under a new PM. They had no manager, no 1:1s, and the seven engineers were split across four completely unrelated projects. They barely felt like a team. So instead of disappearing into the code, the author spent most of the first month talking to people and managing the day-to-day. Ninety percent of their day was in meetings. The perfect plan died on day one. A lesson here is that all those 30-60-90 day plans are nice, but every team has a different situation. The author also learned the hard way about pushing the team too hard. They wanted some early successes, so they pushed to find minimal scope, but the result was everyone always feeling behind and pressured. A veteran engineer took them aside and told them to stop rushing. There are things teams need to learn and improve on, but the manager was being unrealistic. The takeaway is that you need to pay attention, overcorrect, and then listen when someone tells you you've gone too far.

Shifting to developer tools. Roo Code, the open source AI coding tool built for VS Code, is shutting down its VS Code extension and cloud services on May 15th. The team is pivoting to Roomote, a cloud-based coding agent that runs tasks end-to-end across tools like Slack, GitHub, and Linear, producing ready-to-review outputs for developers to inspect and refine. The CEO wrote that their own internal team was already moving away from using Roo Code inside the IDE, instead running it in remote cloud environments where agents could take on multiple tasks in parallel without direct oversight. If the agent can create a good PR from a single prompt, the interaction model changes completely. You let go of the IDE and focus on driving things end-to-end. The agent doesn't just help engineers. It wipes entire types of work off their plate and delivers something nobody has to clean up. It's amazing how fast the consensus shifted. Instead of GitHub at the center of the dev universe, and IDEs as the uniform front door for technologists, neither are true anymore.

Armin Ronacher has a piece titled "Before GitHub" that looks at how we managed source code and packages in the pre-GitHub era. Armin remembers SourceForge, running his own Trac installation with Subversion repositories, tickets, tarballs, and documentation on infrastructure he controlled. He moved to Bitbucket, then eventually to GitHub. He writes that a large part of his Open Source identity formed there. Projects he worked on found users there. Many professional relationships and friendships started because some repository, issue, or pull request made two people aware of each other. That's why he finds what's happening to GitHub today so sad and disappointing. GitHub was part of the social infrastructure of Open Source. But now GitHub is slowly dying. People are tired of the instability, the product churn, the Copilot AI noise, the unclear leadership, and the feeling that the platform is no longer primarily designed for the community that made it valuable. Mitchell Hashimoto announced that Ghostty will move away from GitHub. Strudel and Tenacity moved to Codeberg. Armin argues that if projects move to self-hosted forges, we risk losing things we don't want to lose. The code might be distributed in theory, but the social context often is not. Issues, reviews, design discussions, security advisories, old tarballs, they disappear much more easily than we like to admit. His conclusion is that we absolutely need archives regardless of whether GitHub is here to stay. Something with the power of an endowment or public funding to keep it afloat. Something whose job is not to win the developer productivity market but just to make sure that the most important things we create do not disappear.

On the more technical side, Arcjet published a post on building a production MCP server in Go. They built an MCP server for Arcjet with tools for security briefings, traffic analysis, anomaly detection, request investigation, threat intelligence, and dry-run impact analysis. The first version was a standalone Node.js server. It worked as a prototype, but they quickly rewrote it directly into their existing Go API. That let them reuse session validation, authorization, middleware, tracing, logging, CI, and deployment, while giving tools direct access to their threat detection and analytics systems. One key insight from this post is the shift in mindset from MCP as thin API wrapper to MCP as agent workflow design. The first version just exposed existing API endpoints as tools. Useful but not very interesting. An agent could list your sites, but it couldn't tell you anything you didn't already know. The composite tools are more interesting because they match how people actually think about security operations. Nobody opens a dashboard and thinks I want to GET an analytics endpoint. They think is anything weird happening? Tools like get-anomalies compare a recent period to the previous period and flag traffic spikes, geographic shifts, new bot signatures, and suspicious IP surges. This doesn't map to any single API endpoint. It's a composite analysis that only makes sense as an agent-facing tool.

Moving to productivity and consumer stuff. Gemini can now generate files directly in the chat. PDFs, Microsoft Word and Excel, Google Docs, Sheets, Slides and more. Export your budget proposal to an Excel file, arrange loose ideas into a bulleted draft, consolidate a lengthy collaboration into a single-page PDF. The feature is available to all Gemini app users globally right now. And Google Translate is celebrating twenty years. That's one of those magical services we take for granted now. Visual translation using Lens has gone from a cool trick to a daily travel essential. And translation is one of the top ways people use Circle to Search on Android.

Finally, a piece from the Mooreds blog on why meetings are forcing functions. Many organizations face a common challenge. A complex project that requires effort and perspectives from multiple people, moves through definition and execution phases, and unfolds over weeks, months, or years. But where the tasks to accomplish the project are not anyone's full-time job. Everyone has other obligations, fires to put out, emails to answer. It's easy for long-term strategic work to sink to the bottom of everyone's todo list. One effective solution is a standing meeting. The key is maintaining an agenda and opening each meeting by reviewing the to-dos from the previous one. This creates pressure on everyone to make progress. When people know they'll be asked what's the status of X at the upcoming meeting, it's easier to carve out time for that work amid the daily chaos. This approach works across organizational boundaries too. If you're a consulting firm, a regular cadence of meetings with your client creates gentle but real accountability.

That does it for episode 773. Big themes this week: the hyperscalers are building judgment layers while the market argues about GPU prices, enterprise software incumbents are facing AI-native challengers for the first time, and the way we think about development workflows is changing rapidly. Catch you next time.

260 things we announced at Google Cloud Next '26 – a recap
What Happened When We Treated AI Like an Engineering Teammate (unavailable)
Agents CLI in Agent Platform: create to production in one CLI
Meetings are forcing functions
Generative UI explained without the hype
Managing a team that didn't choose you
1,302 real-world gen AI use cases from the world's leading organizations
Workday's Last Workday?
Building a PCI-DSS Compliant GKE Framework for Financial Institutions
Bed Bath & Beyond CEO: AI will lead to 'significant reduction in headcount.'
The Industry Is Watching GPU Prices. Google Just Moved the Fight to the Judgment Layer
Roo Code pivots to cloud-based agent, says IDEs aren't the future of coding
Before GitHub
Building a production MCP server in Go
You can now easily generate files in Gemini
Celebrating 20 years of Google Translate: Fun facts, tips and new features to try