↵ select ↓ ↑ navigate esc close

Seroter's Daily Reading: Daily Reading List – October 23, 2026 (#895)

Seroter's Daily Reading ·

Listen: https://blossom.buildtall.systems/95657da2c6e0295181ad0d4b526bb6e985b11827dba9cd20384eed7c8dd87c42.mp3

Source: https://seroter.com/2026/10/23/daily-reading-list-october-23-2026-895/


Today's episode covers Daily Reading List – October 23, 2026 (#895). Richard Seroter opens by saying that's a wrap on the week and he's ready for a couple of days off.

First up, Tokens to Deploy: Agent Skills are Great but CLI’s are Better. The article itself did not come through in this feed, but Seroter calls it a great post and says it asks which deployment approach eats the fewest tokens. He notes that Esther looked at the viable options and brought receipts, meaning she compared agent skills and command-line tools with actual measured costs rather than claims.

Next, When are developers receptive to proactive AI suggestions? This is a Research-Driven Engineering Leadership piece describing a five-day field study by researchers at JetBrains, Carnegie Mellon, and TU Delft. They built a proactive code-quality feature into JetBrains Fleet and deployed it with fifteen professional developers. The feature fired after ambiguous prompts, after declined AI edits, and after commits. The big finding: suggestions after a commit drew the highest engagement, at fifty-two percent, while suggestions after a declined AI edit were dismissed sixty-two percent of the time. Proactive suggestions also took less than half the evaluation time of requested ones. The practical advice is to anchor AI review to natural workflow boundaries, break acceptance data down by trigger, and let developers configure when the AI speaks. Seroter says it is pretty clear that developers only want eager AI suggestions at natural boundaries of their work, not randomly in the middle.

Then comes Eight reasons coding agents are better than software developers. This InfoWorld argument insists that coding agents are tireless, faster, ego-free, conversant in every language and framework, happy to write tests and documentation, capable of holding huge context, ready with zero onboarding time, and easy to multiply into a legion. The author says refusing to use coding agents is a career-limiting move. Seroter notes the writer will have the counter piece, why developers are better than agents, next week, so don't freak out. But he adds that if we're being honest with ourselves, these points are valid.

Getting Started with the Google Cloud CLI Remote MCP Server is next. This post was not retrievable, but Seroter's note positions it as the natural extension of the tokens-to-deploy discussion. CLIs are an efficient and effective tool for agents, he says, but maybe your agent can't call local CLIs or tools. That is where a CLI-over-MCP comes in. The title suggests it shows how to expose the Google Cloud CLI through a remote MCP server, giving agents a managed way to run cloud commands without local access.

Then, Rewrite Everything in Java: Not Rust. Shai Almog of Codename One writes that he tested rather than argued. His team built a Java word processor with Codename One that implements the same feature catalog as Wordcraft, an open source Rust word processor. On the same CI runners, the Java version was faster, smaller, and used less memory, with about two-thirds of the code and much cheaper test runs. The measurements include typing latency, startup, RAM, binary size, and CI time; the Java app also builds for phones and browsers. Almog argues Rust is great for kernels, drivers, and database engines, but application rewrites are not the right trade. Seroter says this is a bit of cold water on the Rust-all-the-things movement. Use whatever language you want, he adds, but some are wildly overhyped right now.

Next, AI makes it easy to migrate. Flutter makes it easy to scale. Written by Carlos Mendonça and Emma Twersky on the Flutter blog, this piece takes aim at the idea that because AI can generate Swift and Kotlin quickly, you should just build twice. The authors argue codebases are accumulated memory, parallel agents introduce non-deterministic drift, and verification, not generation, is the real bottleneck. Flutter's hot reload and deterministic widget tree make a better agent harness, and one codebase scales across mobile, web, desktop, and embedded targets. The post also shares adoption stories from Xiaomi, Universal Destinations, Toyota, BMW, and Nubank, and notes that user ratings show users do not evaluate frameworks. Seroter says to read this. It is a good argument for why you should be thoughtful about duplicating codebases for different platforms.

Then, Software is slow because of bad design, not slow languages. Sean Goedecke argues that most programs are I/O-bound rather than CPU-bound, so rewriting them in C or Rust only speeds up a small fraction of the work. Slow languages already delegate heavy computation to fast libraries, especially in the machine learning ecosystem around Python. The meaningful speedups come from redesigns: caching slow operations, doing work asynchronously or in parallel, fixing N+1 queries, and skipping unnecessary work. He suggests using AI to profile programs and find design improvements instead of blindly rewriting. Seroter's take is direct: tell me he's wrong. Your speed problem is not about your choice of language. It is design.

Next is AI harness crash course. This is a YouTube video, and the feed did not carry substantive show notes, but Seroter notes that people are using the word harness to represent many things, and he says Annie does a good job here explaining what a harness is about. The title frames it as a crash course on the term harness.

Then, When to use a diffusion LLM: I compared DiffusionGemma and Gemma 4. This one also did not come through, but Seroter's note explains the trade-off. Diffusion LLMs are cool, he says, and instances like DiffusionGemma generate entire token blocks at once, which is super fast. But he adds that Karl shows regular Gemma 4 is still the best choice for step-by-step reasoning. The title asks the right question: choosing a diffusion model depends on the shape of the workload.

Next, PHP2Shell was only the beginning: AI is accelerating the patch race. This piece from The New Stack, by Matthew Weier O’Phinney, runs alongside coverage of AI finding a critical PHP flaw and the resulting two-month fix period. The title's argument is that PHP2Shell was not a one-off; AI is now digging up flaws faster than maintainers can patch them. Seroter's take is stark. He asks whether this was the worst week ever in cybersecurity, says possibly, and warns it is not going to get better any time soon.

Streaming comes to API Gateway: real-time and AI-ready APIs, without changing your spec is next. Google Cloud announced streaming support in API Gateway in Public Preview. It covers server-sent events, chunked transfer and HTTP response streaming, WebSockets, and gRPC streaming. The point developers will like is that the same API specification and policies apply; streaming is enabled with one flag at gateway creation. The post also introduces deterministic default hostnames that can be predicted before the gateway exists. For AI workloads, this lets API Gateway front model routing and agent traffic with centralized authentication, quotas, and observability. Seroter says here is a great option if you have streaming results from LLMs, agents, or APIs and want a proxy in front.

Finally, Introducing the Elephant-Goldfish Model plugin: scaling design-driven AI software development. This Google open source blog post announces that the internal Elephant-Goldfish Model plugin is now open source under Apache 2.0. It describes a methodology with two roles: the Elephant is a context-rich, long-horizon session that holds architectural vision and design intent, and Goldfish are stateless, zero-context sessions that validate specifications as objective critics. The plugin automates three phases: brainstorming and co-design, drafting a Markdown design document, and the Goldfish Protocol, which includes a comprehension check, antagonistic review, and implementation readiness check. Seroter asks if you have heard of this. It is a way of thinking about AI coding, and he notes that Google has now open sourced an agent plugin you can use too. Seroter embeds that link under the title way of thinking about AI coding. That piece was not retrievable in this feed, but the plugin post identifies it as the source of the Elephants, Goldfish, and the New Golden Age of Software Engineering framing, where design is the new code.

Across the list, a few threads stand out. AI is changing what it means to write, verify, and reason about software, but the arguments keep returning to design, boundaries, and governance. From token costs to patch races, the week's reading suggests the real bottleneck is no longer writing code; it is deciding what code should do and proving it works.

  1. Daily Reading List – October 23, 2026 (#895)
  2. Tokens to Deploy: Agent Skills are Great but CLI’s are Better
  3. When are developers receptive to proactive AI suggestions?
  4. Eight reasons coding agents are better than software developers
  5. Getting Started with the Google Cloud CLI Remote MCP Server
  6. Rewrite Everything in Java: Not Rust
  7. AI makes it easy to migrate. Flutter makes it easy to scale
  8. Software is slow because of bad design, not slow languages
  9. AI harness crash course
  10. When to use a diffusion LLM: I compared DiffusionGemma and Gemma 4
  11. PHP2Shell was only the beginning: AI is accelerating the patch race
  12. Streaming comes to API Gateway: real-time and AI-ready APIs, without changing your spec
  13. Introducing the Elephant-Goldfish Model plugin: scaling design-driven AI software development
  14. way of thinking about AI coding