Seroter's Daily Reading: Daily Reading List – September 30, 2026 (#878) · Seroter's Daily Reading

Listen: https://blossom.buildtall.systems/ecd0933f060f1bb7208824e5c2b23a07e252b004df4c457b1d43887dfb44756a.mp3

Source: https://seroter.com/2026/09/30/daily-reading-list-september-30-2026-878/

Seroter's Daily Reading. Today's post is Richard Seroter's Daily Reading List for September 30, 2026, number 878. Seroter opens with a personal note: he attended the San Diego Padres playoff game against the Chicago Cubs last night, called it the rowdiest, loudest, most fun baseball game he has attended, and then had two major speaking gigs today while very hoarse. As he puts it, he plans well.

First up is Gemini 4 Argon: our next era of frontier intelligence from the Google blog. Seroter's comment is direct: are you not entertained? He says the team shared information about this model, and the numbers look great. The post, by Koray Kavukcuoglu at Google DeepMind, announces the new frontier model. It is built for sustained deep reasoning across long-horizon workflows, with an industry-leading one million token output limit, introductory pricing of two dollars per million input tokens and ten dollars per million output tokens, and strengths in software engineering, legal and finance work, and cybersecurity defense. Argon is rolling out first to trusted cyber defenders through the Fairwind Program. The post highlights internal results at Google, including a quantum optimization example that beat a published baseline by forty percent, memory optimizations freeing over three hundred terabytes, and large-scale C and C++ to Rust migrations, including work on the Fuchsia Zircon kernel. It also cites top scores on DeepSWE, the Vals Index, long video understanding, and CWE-bench for vulnerability remediation. Seroter's playful framing points to the strong benchmark claims.

Next, ZDNET's Joe McKendrick reports on Seroter's own talk in The best AI leaders are ‘a little bit off the wall,’ says Google Cloud executive. Seroter confirms he did say that, and offers it as a summary of some of the points he made in Boston a couple weeks back. The article covers Seroter's appearance at HubSpot's Unbound conference and his message that there is no blueprint for AI. He described what he called chaotic innovation inside Google. A key shift he highlighted is that knowledge and development work used to be about twenty percent thinking and eighty percent execution, but now it has flipped to eighty percent deep thinking and twenty percent execution, which is mentally exhausting. Some top performers are experiencing what he called productivity addiction, wanting just one more problem, and that creates burnout risk. Seroter argued this is not the time for timid leadership, saying the best AI leaders are a little bit crazy, willing to fail constantly, and ready to protect champions. He also warned that when everyone has access to the same tools, companies need to figure out how they stand out, and that prompting well requires real domain knowledge. He compared AI to a crazy intern you have to steer. It is a useful window into how Seroter thinks about leadership in the AI era.

Then Seroter links to Builder.io's Build an agentic software factory, starting with one bug. His take: as you build up a set of AI agents to help you with software, do it incrementally, and this is a useful lesson. The title itself makes the case: starting with a single bug is a manageable way to begin building an agentic software factory, and the post walks through that incremental approach. Seroter highlights it as useful guidance for teams adding AI agents to their software process.

Seroter then highlights Google Cloud's Graph Workflows in ADK: Everything You Need to Know. He says he liked this explanation of working our way up to more sophisticated agent architectures. The post by Annie Wang and Shangjie Chen walks through the Agent Development Kit's Workflow feature using a refund workflow example. It describes graph engineering as breaking a task into nodes, connecting them with edges, and deciding where code, models, or people control the next step. The post covers fan-out and fan-in, deterministic and agent routers, human-in-the-loop pauses, parallel workers, and dynamic orchestration. A key idea is to start with a single agent, then separate independent steps so they can run in parallel, route requests to the right workflow, pause for human review, and process batch cases. It explains when to declare paths in a static graph and when to let Python schedule further work as results arrive. The guidance is to use an edge list when the connections are clear, use dynamic orchestration when results create follow-up work, and keep a single agent for small open-ended tasks. This maps nicely to Seroter's interest in gradually maturing agent architectures.

Next, Seroter points to Vercel's State of agent skills. He notes Vercel sits on a lot of data here thanks to running skills.sh, and suggests reading this for a look at what skills people are installing, for which industries, and how often. The post reports that skills.sh grew to one million agent skills in seven months and recorded nearly two hundred eighty million installs. It compares that pace favorably to GitHub, the App Store, and npm. The supply of skills leans technical, with more than half of listings teaching software engineering, agent workflows, data, infrastructure, or security. Demand is more distributed: software engineering remains the largest category at eighteen percent, followed by agent workflows at fifteen percent, and business operations and writing at nearly eleven percent each. Business operations, writing and documents, and cloud and infrastructure get the most installs per listing, while research and education are oversupplied. Agent workflow skills that improve how agents work are among the most installed. Install concentration is steep: the top three hundred seventy-five skills account for sixty-two percent of installs, while nearly half of all skills are installed exactly once. Seven in eight installs go to cross-industry skills. The report closes by predicting the next million skills will teach company-specific judgment, and that measuring effectiveness will shift from popularity to benchmarks. It is a data-rich look at the agent skills market.

Then there is a16z's State of Markets II. Seroter notes we are getting a bunch of state-of-whatever things lately. The a16z folks call out data about companies using AI and more. David George's post, with over a hundred charts, covers the state of markets for the first half of 2026. It argues that tech is now the everything cycle, contributing about seventy-six percent of S&P 500 earnings growth in 2026, and that the theme of the year is a rotation from bits to atoms, including hardware, semiconductors, power, and networking. It pushes back on GPU obsolescence fears, noting that even older A100s are still renting at or above earlier prices because demand for compute continues to outpace supply. Yet AI adoption remains relatively immature: nearly thirty percent of S&P 500 companies report some quantifiable impact from AI, but only about two percent report a tracked metric. On the consumer side, only about two percent of US households were paying for an AI service as of April. The report also argues there was no SaaSpocalypse, but rather a SaaS-prove-it moment, as software companies traded growth for profitability and the sector repriced. It expects AI to expand the surface area of demand into robotics, biotech, health, and more. Seroter's framing suggests this is one of several state-of-tech reports worth comparing.

Then Google Cloud's Threat Intelligence Group published Vulnerability Discovery and Exploitation Trends in the AI Era. Seroter calls it sobering data about the rise in vulnerabilities and the corresponding increase in exploitation, and notes there are at least suggestions for what to do next. The report finds that vulnerability disclosures doubled in 2026, from around five thousand per month in January to over ten thousand in August. Exploitation nearly doubled from an average of ten and a half per month in 2025 to eighteen per month in 2026. Zero-day exploitation grew only marginally, from eight per month on average to eleven. AI-assisted discovery finds proportionally fewer low-risk vulnerabilities and more medium- and high-risk ones, and fifty percent of AI-discovered vulnerabilities lead to remote code execution, compared with twenty-six percent for the broader ecosystem. The report also covers vulnerabilities targeting the AI stack itself, with agent orchestration frameworks accounting for half of AI-related flaws. It recommends moving from unprioritized mass patching to threat-intelligence-driven triage, targeted edge defense, and automated agentic remediation. Seroter's takeaway is caution with an action path.

Then comes Charlie Guo's Voice Agents Can Just Do Things on Ignorance.ai. Seroter says he buys this more than he did a year ago, but he does not want voice to be the only interface, because sometimes it is better to type things out, or he is in a space where he does not want to be barking commands to a robot. The piece argues that voice agents do not have to talk back. Guo identifies three modes: speech-to-speech, speech-to-action, and event-to-speech. Speech-to-action, where the user talks and the model uses tools, is one of the most underexplored areas, with examples like form filling, creative tools, and computer use. Event-to-speech has the model speak to the user for hands-free or proactive outreach. Guo discusses native audio models like GPT-Realtime and GPT-Live, which preserve tone, emotion, cadence, and backchannels instead of losing them by transcribing first. He encourages developers to expose existing app verbs and nouns as tools, so users can talk to software they already have, and he notes the accessibility benefits. His closing advice is to start with the role of voice in the interaction, not the kind of voice agent. Seroter agrees voice is more capable now, but adds an important note about interface pluralism.

Finally, Google Cloud's Data Agent Kit is now GA: Bring Google Data Cloud to any coding agent. Seroter's summary: do some fairly sophisticated analytics and data science work in whatever AI coding tool you prefer. The post by Arun Nair and Jeff Nelson announces general availability of Data Agent Kit, a free set of Model Context Protocol tools and agent skills that lets the coding agent you already use work directly with Google Cloud data products in Antigravity, Claude Code, Codex, and other popular tools. GA adds support for BigQuery Graph, Bigtable, and Managed Service for Apache Spark, along with quality-of-life improvements. The kit gives agents context about your environment: which tables exist, how they are partitioned, which ones your team trusts, and why last night's job failed. It includes MCP tools for more than fifteen Google Data Cloud services and Google-authored skills for best practices. It works as an IDE extension or a CLI plugin, and it covers analytics, operational databases, the lakehouse, and pipelines. The agent runs with your own IAM permissions, and admins can govern access with VPC Service Controls, Model Armor, and Principal Access Boundary policies. The post says it is included at no additional cost, with standard pricing only for the underlying services.

Across this reading list, Seroter ties together the current AI moment: rapid model progress, maturing agent architectures, a fast-growing skills ecosystem, market data on early adoption, sobering security trends, evolving voice interfaces, and practical data tooling. The throughline is that there is a lot of raw capability arriving, but thoughtful, incremental adoption still matters.

Your signing session ended