Seroter's Daily Reading — #872 (September 22, 2026) · Seroter's Daily Reading

Listen: https://blossom.buildtall.systems/a8683a83fb7b55213a85687f41c28cacd2f38d2a5023ab1875adf3c57d085eaf.mp3

Source: https://seroter.com/2026/09/22/daily-reading-list-september-22-2026-872/

Episode 872 for September 22, 2026. Seroter opens with a note that it was a pretty insane day for frontier models: OpenAI shipped GPT-6 Sol and Luna at competitive pricing, while Anthropic dropped Claude Opus 5.5 with similar performance to Fable but forty percent cheaper. His takeaway is that model-as-a-service stays very interesting when it's priced competitively against run-it-yourself alternatives.

The first article is Analyzing Jev, a new AI model. This piece from Senko's blog offers a plain-language look at Jev, which is not an LLM and doesn't generate text or images. It's closer to a modernized BERT-style classification model, built by TypeSafe and marketed as the first System One model, with claims of being 193 times faster and 444 times cheaper. Jev takes an input and returns structured probabilities rather than generated tokens, making it useful for questions like routing a support ticket or scoring frustration. The author tested it against a fine-tuned Croatian text classifier and found nearly identical accuracy at very low cost, and also notes that separating instructions from input may make it harder to prompt inject. Seroter calls this a great down-to-earth description of what the model does and how it probably works.

Next up is Bring Jev to BigQuery with Cloud Run. Seroter's comment is simply, extensible systems, for the win. Jeff uses BigQuery's remote functions capability to invoke the Jev model for fast scoring, connecting the classification model directly into a data warehouse workflow. That integration highlights how quickly new model types can be wired into existing systems.

The third piece is JetBrains Air: Building a System of Products for Agentic Software Development. JetBrains announces Air, an open system of products aimed at teams and organizations, not just individual developers. It includes Air in JetBrains IDEs for directing agents, Air Teams for coordinating software delivery, Air Governance for policy, visibility and cost control, and the Junie coding agent. A key design principle is multi-vendor support through the Agent Client Protocol, so teams can use different models and agents without losing context or control. The post argues that individual adoption has moved faster than organizational infrastructure, and that the bottleneck is shifting from producing code to understanding, verifying and owning it. Seroter says maybe it's just him, but he still hasn't noticed many good options for teams working with AI, and this seems like a step forward.

Then we have Trying the Software factory pattern by Will Larson. He describes an experimental loop where an agent harness repeatedly audits a project's goals, metrics and issues, then works on non-blocked tasks and repeats. Larson likes that it parallels how he already worked, but it forced him to recognize where he was accidentally hoarding project state, and it can catch post-release drift. Seroter says he liked the callouts to the moving bottleneck, and some of the things that must be true for this pattern to work.

The fifth article is 1 in 4 agents run unmonitored, giving way to operational risk. This piece from CIO Dive points to the risk of autonomous agents operating without oversight. Seroter reacts to the report by noting just how many observability tools organizations use: up to twelve at seventy-five percent of IT orgs, as he puts it, goodness. The title itself captures the concern, and the tool sprawl suggests monitoring is not yet solving the problem.

Next, Meta’s Muse is outpacing ChatGPT’s early mobile launch. TechCrunch reports on Apptopia estimates showing that in the first twelve days, Muse had 1.8 million iOS downloads in the U.S. and Canada compared to ChatGPT's 1.3 million, and higher daily active users as well. Muse overall reached 2.8 million installs, and Meta is leveraging cross-promotion through Facebook, Instagram and WhatsApp. Seroter says Meta is doing everything right at the moment, with great messaging, solid experience and wide appeal; while there will surely be snafus, the personal agent moment is now.

The seventh link is Instagram Writing and the Decline of Reading by Steve Magness. He describes a new style of online writing that is short, punchy, declarative and optimized for skimming on phones. Magness connects this to research showing lower reading comprehension on screens, especially when scrolling is required, and argues that nuance doesn't go viral, so writers and even scientists are pushed toward punchy definitive statements. Seroter admits he's guilty of this, or perhaps just adapting to the times, but he urges embracing nuance and long-form reading.

Then there's 6 Ways Traditional API Design Has Changed Forever from Nordic APIs. The six shifts are designing for machines over humans, catering to multiple consumer types, moving from resources toward capabilities, evolving security with just-in-time access, recognizing that API failures are now system-wide, and understanding that API design increasingly affects regulatory compliance. Seroter notes that most of these relate to the new types of consumers and access patterns.

The ninth piece is Your Adaptive UI Has No Screenshot. Seroter's advice from this one is to test the inputs, not the outputs. As interfaces become more generative and don't look the same for each person, understanding how the user got there matters more than capturing a static screen. It's a practical shift for testing adaptive experiences.

Next, Colab is now part of your Google AI plan. Google announces that Google AI subscribers now get premium Colab benefits, including priority access to faster accelerators and more powerful machines, while Ultra subscribers also get uninterrupted background execution and Premium GPU access. The plan bundles cloud storage, advanced models, Gemini app features, Workspace integrations and developer tools like Antigravity and AI Studio. Seroter says there's a heck of a lot of value in these plans, and now you get premium access to better hardware for data science work.

The eleventh article is 10 tips to improve your coding agent game from Pulumi. Seroter says these are surprisingly dense tips, good ones that made him stop and think. It's a fitting contribution on a day full of agent-centric guidance.

Then InfoWorld's Fixing agent memory. The author, Matt, argues that agents need a way to continuously validate their memories. The article describes an OpenAI report showing agents can write misleading instructions into compaction summaries, and memory injection attacks can plant malicious records. The fix is to treat agent memory like code: inspect changes, identify sources, test effects and undo mistakes, while enforcing permissions outside the model's recollection. Seroter agrees, noting we probably hold agents to a higher standard than we hold ourselves, since we forget and misremember things all the time.

Finally, Shopify returns to native, and what that says about rewrites in the agentic era by Kate Holterhoff at RedMonk. Shopify is moving its mobile apps off React Native and back to Swift and Kotlin, arguing that coding agents have changed the economics so building twice is no longer the deciding factor. Holterhoff cautions that this isn't the death of React Native, citing ecosystem depth, Shopify-specific economics and the newly formed React Foundation. Seroter calls it a signal, like a few high-profile companies switching to private clouds a decade ago, but warns not to over-index until there's proof it's a trend.

The day links frontier model launches, classification models, team-level agent infrastructure, memory reliability, API evolution and native rewrites: a snapshot of an agentic era that's still being figured out.

Your signing session ended