Seroter's Daily Reading — #874 (September 24, 2026) · Seroter's Daily Reading

Listen: https://blossom.buildtall.systems/868b2c2a631418e13fa7ddd0041ecdaa74bfea0db4815030dbd478204e182170.mp3

Source: https://seroter.com/2026/09/24/daily-reading-list-september-24-2026-874/

Episode 874 of Seroter's Daily Reading for September 24, 2026. In his opening note, Richard Seroter reflects on why this newsletter exercise helps him: we all notice a lot of information each day, but taking a moment to write it out by hand, with no AI, imprints those ideas more durably on long-term memory. It's a reminder to stop and reflect on what we're learning.

First up is The Pragmatic Guide to Agent Skills. Seroter says a year ago there were no agent skills, and now there are probably too many. Still, he argues there are credible places to apply them, and this guide from Daniela offers useful instructions for doing that.

The next article, A new, no-compromises database architecture for the agentic era, comes from Google Cloud's blog. Seroter's take is blunt: despite your best efforts, your relational database won't handle a burst of queries from a swarm of agents, and he calls AlloyDB's solution fascinating. The authors argue agentic workloads break the old database tradeoffs between scale, latency, and isolation. Agents are dynamic, unvetted, and bursty, sometimes demanding a thousand compute nodes over a single database within seconds. AlloyDB's architecture is built on three tenets: isolation by design with real-time data access, sub-millisecond baseline I/O, and scale from zero to thousands of nodes. The system uses dedicated production clusters, ephemeral microVM-based agent nodes that read directly from Colossus storage segments, and Google's Jupiter network. In tests, scaling from one to a thousand agent nodes produced near-linear throughput with no measurable impact on the primary database. The goal is to share the data and share nothing else.

Then comes What makes a great co-worker in an AI-native workplace? from Research-Driven Engineering Leadership. Seroter asks what we want from an AI collaborator, and whether we expect the same things from AI as from human teammates; the research suggests the priorities are different. In a study of over fifteen hundred knowledge workers, technical skill ranked second of seventy-five qualities for an AI co-worker but twentieth for a human one. People wanted AI to be competent and legible, including signaling uncertainty, while warmth-related qualities like empathy and caring drew far more objections. Accountability stayed firmly with humans, ranking third for people and fifty-third for AI. There was also a clear handoff etiquette: nearly two-thirds of respondents said reading and verifying AI-supported work is the minimum owed before it changes hands.

Google Cloud: Investing in the future of PostgreSQL — 2026 highlights is next. Seroter admits he didn't realize Google Cloud contributed so many things to open source PostgreSQL, and calls it a good roundup. The post summarizes technical work between January and September of 2026. It covers logical replication and conflict management, including conflict log table infrastructure planned for PostgreSQL 20; core engine performance improvements such as faster sequential scans and more accurate promotion timeouts; catalog, collation, and indexing bug fixes; and hardening of extensions like plpgsql_check, pgfincore, and pgtt. The piece also highlights contributors including Dilip Kumar, Nitin Motiani, Robert Pang, Aniket Jha, and Sunaina Punyani.

How to Run Open Models in the Cloud Without Going Broke could not be retrieved, but Seroter's commentary gives the core advice: scale to zero, use spot VMs, and right-size your hardware. He describes it as simple but relevant guidance.

Our Favorite Management Tips on Building Influence, from Harvard Business Review, compiles ten favorite tips from HBR's Management Tip of the Day newsletter on the theme of influence. Seroter says these are good tips, and he emphasizes that building influence is an important area to focus on if you want to drive people toward a desired action.

Introducing Support for Local AI Models in the Antigravity SDK, from the Google Developers Blog, announces that the Antigravity SDK now supports local workflows across a range of local models, with initial support for Gemma 4 26B A4B using Google AI Edge's LiteRT. Seroter says he's glad to see this, and notes you can use the SDK to build with an AI harness using local models, with many benefits. The post highlights cost efficiency without API charges or rate limits, privacy by keeping code and requests on the local machine, offline resiliency, and hybrid workflows that combine local and cloud models. A recorded hybrid example used a cloud model as the architect and local Gemma instances as the workforce, with over ninety-seven percent of tokens running locally while keeping proprietary code on-device. The SDK also supports OpenAI-compatible local servers such as Ollama, LM Studio, or vLLM.

Will TypeSafe’s Jev Change How We Build AI Applications? by Laurie Voss on O'Reilly Radar asks whether a new kind of model reshapes AI application architecture. Seroter wonders if a bunch of business logic was overkill for large language models but makes sense for classifiers like this, and suggests the shape of your code may change again. Jev, from TypeSafe, classifies, scores, and routes but doesn't generate free-form text. It returns typed answers with probability distributions, and early numbers claim it's up to two hundred times faster and four hundred times cheaper than a general-purpose model for decision tasks. The tradeoff is that it doesn't explain its reasoning, so Voss suggests pairing cheap Jev decisions across every trace with sampled LLM judges for directional signal. He argues that trustworthy probability scores could let teams set thresholds and escalation paths more reliably, and that cheap decisions change the math of how we build and measure AI applications.

Behind Project Suncatcher, our moonshot to put AI in space, from Google's blog, covers a real effort to test AI hardware in orbit. Seroter's reaction is simply: how wild is this, and it's real. The project is launching a prototype satellite on the SpaceX Transporter-18 rideshare mission, developed with Planet, to see how Google TPUs survive space conditions. The team is tackling vibration and g-forces during launch, radiation effects on chips, and cooling without airflow in a vacuum, likely through heat pipes and radiators. Future plans include clusters of satellites linked by high-bandwidth lasers, with a two-satellite interconnectivity test planned for 2027. This first mission is about learning what works before scaling up.

Do Frameworks Matter Anymore? by Brooks Lybrand takes on the question of whether web frameworks still have value in an age of vibe coding and agentic programming. Seroter calls it a feisty question, and says you could make a viable argument that we don't need the structure and bootstrapping help of a web framework, but this post argues there is durable value even with AI coding at our fingertips. Lybrand defines frameworks as providers of abstractions, structure, and constraints. He points out that if frameworks truly didn't matter, you'd have no reason to default to React; and if React does give agents an advantage, then frameworks do matter. His larger claim is that everyone is either using an explicit framework or generating an implicit one: an LLM asked to build a website will create its own abstractions. He still wants code that is well tested, secure, easy to reason about, and that gets his agent into the grooves it needs to produce good results.

Finally, What happens when you let an agent design a new git system? by Jason Davenport in The Davenporter describes a project called cosm. Seroter says Jason wondered what type of solution an agent would create for itself, and the answer is a graph that handles agentic workers better than git. The idea is that git struggles with the rate of agentic changes, token costs from reading large codebases, and lineage tracking. Cosm flips the order: the graph and AST representation become the primary source of truth, and the file system becomes a secondary rendered output. Agents can lock specific entities, check the blast radius of a change, and stack updates. Davenport notes real tradeoffs, including difficult test creation, backward compatibility challenges, undefined human-agent interaction, and imperfect AST renderers. His agent learnings include defining principles in an AGENTS.md file early and using subagents for more detailed planning.

Across the day, there's a clear throughline: the agentic era is reshaping not just applications but the infrastructure beneath them, from databases and source control to local model runtimes and even satellites. Meanwhile, the human side still matters, whether that's building influence, deciding how AI co-workers should behave, or remembering to reflect on what we learn.

Your signing session ended