Source: https://seroter.com/2026/10/29/daily-reading-list-october-29-2026-899/
Welcome to Seroter's Daily Reading. Today's episode covers the Daily Reading List for October 29, 2026, issue number 899. Richard Seroter notes that the day started with a 6 a.m. customer session, moved to the GitHub Universe main stage for a panel discussion, follow-on Q&A, and a press interview, and then a flight back home.
First up is Designing Agentic Interfaces with A2UI and Genkit, by Daniela Petruzalek. Seroter says he digs generative UIs right now, and Daniela puts A2UI through its paces with Go. The article walks through upgrading a diagnostic agent called AIDA from a static chatbot to a generative cockpit. A2UI is an open protocol from Google that lets agents send native UI components instead of plain text. The author defines a custom catalog of seven components, including HUD gauges, 3D process graphs, filterable tables, and action buttons. On the Go backend, Genkit middleware validates the agent's A2UI output against that catalog. The frontend renders the stream and sends user clicks back as agent actions. The piece also shows human-in-the-loop tool interrupts: when the agent proposes a shell command, the turn pauses for explicit approval or denial before running anything. Seroter's interest in generative UIs is clear, and this article demonstrates a full production-style implementation.
Next is What happens when the agent is long gone, but its access key lives on? from The New Stack. Seroter admits he hadn't really thought about this problem, but says it seems important to have a strategy for it. The article itself wasn't fully retrieved in the feed, but the title points to a growing concern: AI agents are given credentials and access keys, and when those agents are decommissioned or abandoned, those credentials may linger. Seroter's take is that governance around agent credentials is something teams need to plan for, not discover later.
Behind the scenes, two months later: How we ship Google Agent Skills like products, by Remigiusz Samborski on the Google Cloud blog. Seroter says one thing about Google is that they look at everything as an engineering problem, and for agent skills his team built a robust strategy and set of platforms to generate, support, and eval them. The article describes shipping skills like products, not documents. Every change goes through automated CI gates: structural conformance, linters, semantic versioning, evaluation suite validation, link checks, an AI review agent, and human approval. The AI reviewer catches anti-patterns like the deterministic trap, where rigid CLI sequences are written directly into instructions instead of being moved to scripts. It also enforces a three-tier architecture: concise metadata, compact core instructions, and reference docs. Lessons include making evaluation expectations atomic, gating expensive LLM checks behind fast deterministic ones, and preventing documentation drift with generated indexes. Seroter's commentary highlights the rigorous engineering approach behind Google's skills repository.
The next piece is You can’t find 10x developers, but you can build 10x organizations, by Justin Reock and Scott Carey from the DX newsletter. Seroter says there's no doubt many individuals are now remarkably productive, but software usually happens in teams, and activating the group and its workflow is how you win. The article revisits the Coding War Games study from the 1980s, which found an 11.1x spread in performance, but the spread clustered by company rather than individual talent. Top performers had quiet private workspaces, fewer interruptions, and more flow time. The authors argue that organizational factors like protected focus time, a fast inner loop, and psychological safety are what create high-performing teams. Seroter's commentary reinforces the point that team environment matters more than hiring mythical individual stars.
Coding agents need guidance, not faith from BCS is next. The article itself couldn't be retrieved, but Seroter says good input, relevant context, and smart guardrails are necessary; without that, you're just fooling around. So the core message is about treating coding agents with structured guidance rather than blind trust.
Then there's How not to run an AI code review. Serve the elephant whole. by e5bravo. Seroter says we're all figuring it out as we go along, and these are some important lessons. The author describes a personal experiment: seven AI models were asked to read a 111,000-line codebase and produce eleven reports. All seven claimed they were done, but none actually were. One claimed full coverage while missing most files. The mistakes were design flaws: serving the whole codebase at once, never defining what done meant, letting workers report their own progress, and not budgeting for checking. The fix was to split the work into small units, write a strict definition of done, compute the scoreboard with a script, and use a cross-family check. The final results showed recall between 12% and 37% for different models. Seroter's point is that these lessons matter because AI code review is still immature, and process design determines whether you get useful results.
Gemini Robotics Embodied Reasoning 2 is now available on Google Cloud, from the Google Cloud blog. Seroter says to use this unique model with the protections and management of Google Cloud supporting it. The announcement covers a specialized Gemini model that serves as a robot's high-level brain. It combines spatial and temporal video understanding with long-horizon planning and built-in safety guardrails. Developers can declare low-level control interfaces as tools and stream multimodal data into the model. It supports task verification, self-correction, multi-robot collaboration, and natural language interaction. On Google Cloud, teams can simulate, evaluate, and deploy physical AI with enterprise IAM and lifecycle management. Seroter's comment positions this as an enterprise-ready robotics model.
Getting good at AI takes time and practice, by Sean Goedecke. Seroter calls this nothing earth-shattering but a helpful reminder that using AI is a skill of its own. The article argues against the idea that current models need no learning curve. Skilled AI users learn to supply the right context, know what the model can and can't do, recognize the specific mistakes AI-generated code tends to make, and quickly skim through AI slop to find the useful claims. These skills may last through the centaur age, where human-machine pairings outperform either alone. Seroter's take is straightforward: practice with AI is essential.
You can now generate, preview, and edit documents, PDFs, presentations, spreadsheets, and more directly in Gemini, from the Google blog. Seroter says they're late in adding this, but he's happy it's there now. The announcement describes new Gemini app features for creating and refining documents, PDFs, presentations, spreadsheets, and code directly within a chat. Users can expand to full screen, edit directly, highlight text to ask questions, or have Gemini refine files. PDF preview and editing works on web and mobile, while other formats are web-only. The example shows turning handwritten lecture notes into an editable PDF study guide.
AI agent access is moving faster than identity controls can govern from VentureBeat is next. The article itself couldn't be retrieved. Seroter says he's seeing an uptick in this type of content, which is probably good because the industry needs to figure out what production-grade agents look like. The title suggests a growing gap between how quickly agents are granted access and how slowly identity systems adapt to govern them.
Design Tenets for Agent Identity Management Systems from NIST, by Harold Booth. Seroter notes he hasn't seen a lot of NIST content lately, but this is useful and timely. He suggests checking it out for a checklist of what you should be doing to properly manage agent identities. The article's body was not included in the feed, but the title and Seroter's comment point to concrete design principles for managing identity and authorization for software agents.
Finally, Near-zero downtime upgrades: Announcing Cloud SQL for MySQL blue-green deployments, from the Google Cloud blog. Seroter says this takes some risk out of database upgrades. He likes that there are no app changes and that it's automated with near-zero downtime. The announcement describes a fully managed blue-green deployment process: Cloud SQL creates a green clone, continuously replicates data, and then performs a switchover. On Enterprise Plus, cutover downtime is sub-second. The existing IP address, instance name, and backups transfer automatically, so developers don't change anything. Teams can test against real production data before cutting over. Seroter's summary captures the value: lower risk, less effort, and near-zero downtime.
That wraps up today's list. A few themes stand out: agent identity and credential governance keep coming up, generative UIs and agent skills are maturing fast, and team-level factors still dominate individual productivity. Thanks for listening.
- Daily Reading List – October 29, 2026 (#899)
- Designing Agentic Interfaces with A2UI and Genkit
- What happens when the agent is long gone, but its access key lives on?
- Behind the scenes, two months later: How we ship Google Agent Skills like products
- You can’t find 10x developers, but you can build 10x organizations
- Coding agents need guidance, not faith
- How not to run an AI code review. Serve the elephant whole.
- Gemini Robotics Embodied Reasoning 2 is now available on Google Cloud
- Getting good at AI takes time and practice
- You can now generate, preview, and edit documents, PDFs, presentations, spreadsheets, and more directly in Gemini
- AI agent access is moving faster than identity controls can govern
- Design Tenets for Agent Identity Management Systems
- Near-zero downtime upgrades: Announcing Cloud SQL for MySQL blue-green deployments

