Listen: https://serve.buildtall.systems/ca2b1e2c2bf1ad5013c959763a4788efd05d464c07726396d95480847d03a1a2.mp3
Source: https://seroter.com/2026/11/16/daily-reading-list-november-16-2026-910/
Seroter's Daily Reading: Daily Reading List – November 16, 2026 (#910). Richard Seroter opens by saying it's good to be back home after a fun and exhausting week in Japan, and that the flights were touch and go due to the temporary shutdown. He notes today's list features some very opinionated advice.
First up is Can AI fix its own slop? I tested 7 “humanizer” skills. Seroter says it's getting a bit tougher, but most of us can spot AI-generated content. He adds that Karl benchmarked seven skills that remove signs of AI writing. The title suggests the piece tests whether AI can clean up its own generated text, though the full article was not retrievable.
Next, Designing an Engineering Strategy. Seroter observes that most of what people call strategy is really a vague vision or a set of tactics. This post by Aleix Morgadas recounts good descriptions of how to discover the core problem to solve. Morgadas draws on Richard Rumelt's definition: a strategy is a coherent set of analyses, concepts, policies, arguments, and actions that respond to a high-stakes challenge. He adapts that to engineering, arguing that an engineering strategy must address a high-stakes business problem from an engineering perspective. The article offers questions to find such problems, distinguishes strategy from vision and fluff, and warns that bad strategy often comes from fluff, failure to face the challenge, mistaking goals for strategy, or bad objectives.
Catch Go bugs that fuzzing misses with Gemini and a constraint solver. Seroter's point: when you combine an LLM with a constraint solver, you can uncover subtle bugs in Go code that unit tests and fuzzers miss. Karl Weinmeister from Google describes compiling Go functions to RISC-V assembly, identifying surviving bounds-check traps, and having Gemini 4 Argon translate the paths into Z3 constraints. The solver finds concrete crashing inputs, which Go test then confirms. The approach caught planted bugs and latent issues in skill-lint, while fuzzing missed two because Go's fuzzer rarely exercises slices with no spare capacity. The fix: clip slice capacity to length.
Subagents considered harmful? Seroter says the author pitches smart forks with relevant context versus subagents working from summaries. The article by Nik Vdp argues that summaries are lossy and subagents often miss critical context. He describes a failure where a subagent missed the main question. He compares fresh subagents with forked subagents that inherit full context, then proposes smart forks that use Jev-like system one models to filter which context to keep. He concludes that delegation is a theory-of-mind problem and that smart forks may offer the best of both.
Advancing science & engineering with Antigravity’s /teamwork. Seroter calls this post amazing, and recommends reading real-life examples of agent teams solving problems at an unheard-of scale. Vahab Mirrokni describes how Antigravity's /teamwork pairs parallel exploration with verifiers and partitions context. The examples include proving the Courtade–Kumar conjecture and formalizing it in Lean with 50 million lines of code, formalizing the Weil conjectures, Hilbert's tenth problem, and parts of Navier–Stokes; achieving 95% top-1 accuracy on NMR molecule structure; building FarArray for trillion-edge graphs; rewriting Zircon in Rust; optimizing TPU kernels; and training a Go engine from scratch to 1,885 Elo.
Your coding agent burns two-thirds of its tokens after the first edit. Seroter notes that agents chew a lot of tokens in post-edit loops of re-reading and testing, and that making them declare a plan before editing can cut that waste drastically. Andreas Rau covers a paper called FlowPlan. The paper finds that in Claude Code, about two-thirds of tokens are spent after the first edit across roughly 31 further calls. FlowPlan asks the agent to write a structural plan as graph operations before editing, then verifies the diff against the plan. This drops the post-edit tail from about 31 calls to 12, cutting token use significantly, though resolution doesn't change statistically. The author cautions that the plan itself isn't checked, and that a wrong plan faithfully executed can pass. He predicts declaring intent then verifying against it will become standard.
Goodhart’s Law and AI agents: how to measure developer productivity now. Seroter's take: measuring output doesn't matter much now; measure your delivery system and outcomes, but don't obsess over any one number. Pere Pages explains that when agents make output nearly free, metrics like pull requests and lines of code become easy to game. He reviews research including DORA, SPACE, and DX Core 4, and recommends measuring team-level speed and stability together, using anonymous developer surveys, and focusing on user and business outcomes. He also warns that agents can game tests and that perceived speed can be wrong.
Google named the furthest in vision for the Gartner Magic Quadrant for Data Lakehouse Platforms. Seroter says this is the first Magic Quadrant from Gartner for lakehouses, and that lakehouses are becoming a big part of a modern data stack, and Google has a good one. The Google Cloud blog by Sirish Chandrasekaran and Gaurav Saxena announces Google is a Leader and furthest in vision. The post describes the borderless Lakehouse as a foundation for agentic workloads, built on Apache Iceberg, with cross-cloud connectivity, Knowledge Catalog for agent grounding, and examples from Commerzbank, Bloomberg Media, and PayPal. It also highlights performance improvements in BigQuery and Spark.
MCP vs CLI vs Skills: What Each Costs in Context. Seroter says the OpenRouter team makes the point clear that sending every tool definition on every turn is very wasteful, and to read the post for data and advice. The OpenRouter post measures token costs on a GitHub task. A CLI used 1,421 tokens, a skill 3,315, and the default MCP server 59,417 because 46 tool definitions added 19,436 tokens. Tool Search can defer loading and reduce upfront cost, but results still need trimming. The recommendation: use a CLI when possible, scope MCP tools, and cap results.
Introducing A2UI 1.0: What’s New in Google’s Framework-Agnostic Generative UI Protocol. Seroter says the CopilotKit crew wrote up a great summary of the new standard for generative UIs and what you should know. Anmol Baranwal and Eli Berman explain that A2UI lets agents describe UI as JSON referencing a catalog of app components. Version 1.0 adds bidirectional function calls, mixing multiple catalogs, composition rules, one-message surface creation, richer validation, user-initiated functions, accessibility, reserved @ prefix, and extension metadata. CopilotKit's Generative UI is built on it and offers a standard catalog of 70 components.
Go Turns 17. Seroter's commentary: great language, great community, both thriving because of a commitment to outstanding work. Austin Clements, for the Go team, celebrates 17 years since Go's open source release. The post covers Go 1.26 and 1.27: generic methods, new JSON v2 package, goroutine leak profiler, HTTP connection reuse, Green Tea garbage collector, SIMD packages, stack allocation improvements, post-quantum crypto, go fix modernizers, source-level inliner, gopls enhancements, and a new pkg.go.dev API. It ends with looking ahead to Go 1.28.
Pay for Hugging Face Hub with your Google Cloud Account. Seroter explains that instead of Hugging Face charges on a credit card, you can latch them onto your existing Cloud bill. The Hugging Face blog by Zaïd Alhadad and others walks through linking a Hugging Face organization to Google Cloud Marketplace. It covers one bill, no new vendor, and everything covered. The process involves subscribing on the marketplace, linking the organization, and approval. Included credits still apply and usage monitoring is available.
The Jev Revolution: What Engineers Are Already Building with System One Models. Seroter says Jev captured an important moment as people suddenly realized they didn’t need reasoning for everything. Latent Space covers the AI Engineer Code Summit, where Diogo Almeida of TypeSafe AI argued that Jev and similar system one models offer reliable intelligence for automation. Examples include Glean using Jev for zero-shot flexibility, UiPath running ranking tournaments of document chunks in under a second, and Workday prototyping predictive UI. Almeida compared the moment to the early internet.
Gemini Decisions API — Fast, Calibrated Multimodal Decisions. Seroter's comment: our new decision model has some useful functionality that seems unique. The article itself could not be retrieved, but the title suggests Google's new Decisions API is designed for fast, calibrated decisions across multimodal inputs. Seroter's note highlights unique functionality, possibly related to returning probabilities and confidence for closed-set decisions.
Benchmarking the Gemini Decisions API. Seroter says Bill goes deep on the Decisions API and shows some solid performance. The author, Bill, on Dev.to, builds a customer-support ticket triage demo and benchmarks the Decisions API against Gemini 3.8 Flash, Gemini 4 Hydrogen, Gemini 3.5 Flash-Lite, and Gemma 4 E2B. The Decisions API scored 46/48 on team routing, 47/48 on refunds, and median 202 ms per ticket, 11.7 times faster than 3.8 Flash. It handled 100 questions in 359 ms. Its confidence score separated right from wrong. The article also builds an ADK agent with the Decisions API as a guardrail and triage tool.
Across the list, several threads stand out: the push to make agents more reliable and less wasteful, from token costs to planning and decision models; the maturation of generative UI and lakehouse platforms; and the steady evolution of Go. Seroter's reading list for November 16, 2026 offers a snapshot of an industry simultaneously chasing scale and grounding.
- Daily Reading List – November 16, 2026 (#910)
- Can AI fix its own slop? I tested 7 “humanizer” skills
- Designing an Engineering Strategy
- Catch Go bugs that fuzzing misses with Gemini and a constraint solver
- Subagents considered harmful?
- Advancing science & engineering with Antigravity’s /teamwork
- Your coding agent burns two-thirds of its tokens after the first edit
- Goodhart’s Law and AI agents: how to measure developer productivity now
- Google named the furthest in vision for the Gartner Magic Quadrant for Data Lakehouse Platforms
- MCP vs CLI vs Skills: What Each Costs in Context
- Introducing A2UI 1.0: What’s New in Google’s Framework-Agnostic Generative UI Protocol
- Go Turns 17
- Pay for Hugging Face Hub with your Google Cloud Account
- The Jev Revolution: What Engineers Are Already Building with System One Models
- Gemini Decisions API — Fast, Calibrated Multimodal Decisions
- Benchmarking the Gemini Decisions API

