Seroter's Daily Reading — #873 (September 23, 2026) · Seroter's Daily Reading

Listen: https://blossom.buildtall.systems/f3151d4160c6e3c79f38d088328a5936375b5d4e939ad33c8309b00f020ac1a3.mp3

Source: https://seroter.com/2026/09/23/daily-reading-list-september-23-2026-873/

Episode 873 of Seroter's Daily Reading for September 23, 2026. Seroter opens by saying it was one of those days that feels productive but had so few pauses that he needs to replay the highlights on the drive home. That is exactly what this episode does.

First is Launching MCP Apps in Toolbox, from the Google Cloud publication on Medium. The article could not be retrieved, but the concept in the title and Seroter's note is Model Context Protocol apps that return interactive UI components along with their data, rather than just text or structured results. Seroter calls MCP Apps an interesting idea, and says it is reasonable if your client can render a web view.

Next is Gemini 3.8 text-to-speech says hello, written by Leland Rechis and Alan Cowen on the Google blog. Google is introducing two text-to-speech models. Gemini 3.8 Flash TTS is built for deep creative direction and character design: create entirely new voices from scratch using natural language prompts, replicate a voice from a thirty-second sample, and direct delivery line by line with control over pacing, emotion, dialect shifts, and even conversational backchanneling such as laughs and sighs. Gemini 3.8 Flash-Lite TTS is aimed at high-volume, cost-efficient uses like dubbing, audio content creation, and expressive voice agents. Both models support more than a hundred languages and dialects. On the safety side, Google includes consent verification for voice replication, SynthID watermarking, and C2PA credentials. Seroter calls these wild models and notes that Simon Willison already built a fun example with them.

Third is Leave the Class Path in the Rearview Mirror, from the Netflix Technology Blog. The article itself did not come through, but the title and Seroter's commentary make the intent clear. Netflix is trying to make Java development easier with better modularity, meaning a move away from the traditional class path. Seroter's response is simply yes please.

Fourth is Keep post-paid Gemini billing: use Vertex, not AI Studio Prepay, by JK Gunnink on Substack. The article describes how Google AI Studio is pushing developer API accounts toward prepaid credits, which expire, are non-refundable, and stop serving API keys when the balance hits zero. Gunnink's workaround is to call Gemini through Vertex AI on Google Cloud instead, tied to a normal Cloud Billing account, so usage accrues on a post-paid cycle. The same Google GenAI SDK works for both paths; you point the client at Vertex with a project and location and rely on application default credentials instead of an API key. There are gotchas around regional model availability, per-project quotas, and avoiding a half-migration that leaves old keys draining prepaid credits. Seroter's take is that vendor efforts to prevent runaway bills now mean you have another thing to worry about: making sure your app keeps running when you have eaten all your credits.

Fifth is The Internet Is Not Ready for the Agentic Wave, from Builder. Seroter agrees with the premise. His summary is that the web assumes you're human. Page layouts, CAPTCHAs, carefully crafted calls-to-action on videos, none of that matters to agents. The article argues the web's assumptions need to catch up with a wave of non-human users.

Sixth is GKE becomes more elastic: Scale to zero, save costs, and keep workloads responsive, by Eyal Yablonka and Scott Funkenhauser on the Google Cloud blog. Google Kubernetes Engine version 1.37 now has a native way to scale workloads to zero and back. It builds on Horizontal Pod Autoscaling with AutoscalingMetric and Kubernetes enhancement proposal 2021, and it can read external metrics directly from Google Managed Service for Prometheus, so teams don't need to run KEDA as an extra component. To handle cold starts, GKE offers capacity buffers, pooled warm compute that lets a pod claim resources immediately instead of waiting for a node. Seroter notes that if you thought the optional KEDA extension was the only way to get a Kubernetes cluster to scale workloads to zero, you're in luck.

Seventh is Building generative media agents: model, harness, tools, from the Google Cloud publication on Medium. The article could not be retrieved, but Seroter describes it as advice for people building agents that produce images, video, and audio. The title itself lays out the architecture: a model, a harness around it, and the tools it can call.

Eighth is The Death of Apps Has Already Begun, from Webdesigner Depot. The argument is that apps will not disappear, but you will stop using them yourself, as agents and assistants handle the direct interaction. Seroter says he is one hundred percent down with this point of view.

Ninth is My comment on MCP was always a bad idea?, from Simon Willison's Weblog. Simon is responding to a Hacker News post that questioned whether Model Context Protocol is always a bad idea. He argues that a full-blown terminal agent with unfettered internet access can BOLO: Monilike. You want control over which external services the agent can access, a way to handle authentication without handing API keys directly to the agent, a sensible UI for connecting more services, and strong audit logging. MCP makes all of that easier to provide. Seroter says this makes sense. And adds that MCP isn't necessary if your client has full access to CLIs and APIs, but many clients don't and won't.

Tenth is How to build a Jev-style classifier with DiffusionGemma and vLLM, from the Google Cloud publication on Medium. The article itself did not come through, but Seroter calls it a cool by Karl. The lesson is that with DiffusionGemma and vLLM, you already have an open, high-performing option for building a Jev-style classifier.

Eleventh is Anthropic made Opus 5.5 cheaper. Then it broke four things your agent depends on, by Amanda Caswell at The New Stack. Anthropic reduced the price of Opus 5.5, but the piece details four things that broke for agents in the process. Seroter broadens the lesson: this applies to any provider. Don't just switch over to new versions. Make sure you understand what got toggled on or off in a given release before your agents depend on it.

Finally, Google Open-Sources AX a Kubernetes Style Orchestrator for Autonomous AI Agents, by Olimpiu Pop at InfoQ. Google has introduced AX, an open-source Apache 2.0 licensed orchestrator and declarative runtime. It runs on Agent Substrate and treats agents as stateful actors rather than microservices or batch jobs. It can checkpoint and suspend idle agents and resume them in under a second. The control plane exposes four Kubernetes-style primitives: Task, Workspace, Gateway, and Model. Task defines execution lifecycle and resource constraints; Workspace assembles environments with Git repositories, MCP servers, and skill bundles; Gateway manages outbound network policies and credential injection; and Model is the unified control point for LLM provider settings and secrets. Developers interact through the ax command line tool, and there is a split in the community between people who like the cost savings and people who see the more pre-building, and what the latter should be. Serote's take is that you can make existing infrastructure agent-ready with the frameworks and orchestrators Google has been open-sourcing lately.

That closes the day. Across these twelve links, the same themes keep appearing: agent-native infrastructure, expressive models with better safety guardrails, and the slow realization that apps, the web, and billing systems all have to change for non-human users. Seriter's drive home has plenty to replay.

Your signing session ended