Seroter's Daily Reading — #876 (September 28, 2026) · Seroter's Daily Reading

Listen: https://blossom.buildtall.systems/9e5e881fd5dc3cddb477e8fab3b5ffdf4b8a3d259214a17a73fbc1d940409b77.mp3

Source: https://seroter.com/2026/09/28/daily-reading-list-september-28-2026-876/

Episode 876, September 28, 2026. Richard Seroter had a relaxing birthday weekend and was ready to go, and as he notes, this reading list includes a handful of AI-related items.

The first article is Experts Lead Experts from SVPG. Seroter's take is that while the principle seems obvious, not everyone believes that experts should lead experts, and some companies put people-only managers in place. The piece argues that the recent trend toward fewer managers and more individual contributors is temporary. At top product companies, skilled leaders who are themselves experts in engineering, product, or design guide their teams. AI has disrupted how products are built, but that makes strong leadership more important, not less. Some companies that cut managers have already started reversing those changes.

Next is Do Angular apps need WebMCP? by Erik Lieben. Seroter points out that an AI agent could figure out a web page through the DOM, but as this test shows, that consumes more time and tokens. The author built a small Angular todo app and had six different models complete tasks with and without WebMCP tools. Every run succeeded either way, but with WebMCP, agents used fewer browser commands in all twelve pairs and were faster in ten, while also using fewer input tokens in most cases. The takeaway is that accessibility is the floor, but well-designed tools add efficiency.

The third item is Elevating Antigravity agent skills, Part 4: Subagent messaging from Google Cloud on Medium. The full text could not be retrieved, but Seroter's commentary highlights the challenge: subagents need to know about each other and communicate. Maybe you used a shared document or some other way to pass messages. Antigravity has some built-in options for this.

Then there is Nine unlikely trends shaping software development from InfoWorld. Seroter found the title fair and noted some unexpected industry movement. The nine trends include plain JavaScript beating TypeScript through type stripping, SQL making a comeback over ORMs, local IDEs winning over cloud dev environments, monoliths beating microservices, happy paths beating integration, on-prem metals beating cloud, specialized engineering beating full-stack developers, Wasm beating Docker, and Java beating Node, Go, and Rust thanks to virtual threads. The common thread is a shift back toward simpler, more direct tools.

Run decision models on vLLM and Red Hat AI using DiffusionGemma comes from Red Hat Developer. Seroter says Red Hat is showing how DiffusionGemma makes a pretty good Jev-style decision model option. Jev, from TypeSafe AI, is a model that returns typed, probabilistic decisions instead of chat. The article explains how the vLLM community added structured-read mode to DiffusionGemma, letting it answer multiple-choice, yes-or-no, and scored questions in a single denoising step. This gives an open, self-hostable alternative to hosted decision APIs, with Red Hat AI validation and preview images.

Meta announces enterprise AI platform, recruits MongoDB CEO to lead it is from VentureBeat. The article text failed to fetch, but Seroter's advice is to keep an eye on this one. Will enterprises buy core technology from Meta? He says don't rule it out. The title itself suggests a significant move: Meta is launching an enterprise AI platform and bringing in the MongoDB CEO to lead it.

Automating coherent long-form video generation is a Google research blog post. Seroter's comment is that today is the least realistic and consistent that AI-driven video generation will ever be, so it only gets better from here. Google introduces a unified multi-agent framework that generates temporally consistent long-form video narratives. It includes the AI video co-director for orchestration, CANVAS for visual storyboarding, A²RD for scaling temporal dynamics, and VQQA for closed-loop refinement. Together they overcome identity drift and cascading failures of earlier linear pipelines.

The next piece is Unlock 3x QPS and microsecond latency with Memorystore for Valkey 9.1 from the Google Cloud blog. Seroter emphasizes that Valkey is not just a Redis alternative; by itself, it is a top performing database for scaled access. Version 9.1 introduces a lock-free multi-queue messaging architecture that eliminates cross-thread CPU waste and achieves up to 3x QPS at microsecond latency compared to Memorystore for Redis Cluster. New commands like CLUSTERSCAN, HGETDEL, and MSETEX, plus database-level ACLs, make it attractive for high-throughput AI and microservices workloads. Major League Baseball and Target are cited as customers.

Apps, Agents, and Aggregation from Stratechery tackles the shift from apps to agents. Seroter says this is all feeding his confirmation bias, but the trend is hard to ignore: we'll use super-apps that aggregate through agents, and generative, personalized UIs for everything else. The article argues that agents are not just AI but AI with access to a computer, and that generative UI is here. Meta's Muse and Microsoft's Copilot are making bids to be the only interface you need, with a computer on your behalf. The scarce resource becomes inspiration, not discovery.

Finally, Why your startup needs open models alongside frontier APIs from the Google Cloud blog. Seroter calls it a good post outlining workloads where an open model is a good complement to frontier model use. The author argues that most production requests don't need a frontier generalist, and a compound AI stack pairing frontier models with compact open-weight models like Gemma 4 wins on latency, cost, and control. Gemma fits edge execution, high-throughput routing, task-specific fine-tuning, and vertical starting lines like MedGemma. Examples include Cue, HubX, K-Dense, and Latitude.

Across today's list, a throughline emerges: AI is becoming more capable as an agent, more efficient through specialized models and tools, and more integrated into how software and databases are built and operated.

Your signing session ended