Source: https://seroter.com/2026/10/30/daily-reading-list-october-30-2026-900/
Welcome to Seroter's Daily Reading. Today's post is Daily Reading List – October 30, 2026 (#900). Richard Seroter opens his nine hundredth issue by noting he likes big round numbers, and he tells readers in America to enjoy Halloween and keep the best candy for themselves.
First up is Antigravity Plugin Marketplace in Action: Auditing My Talk Slides with the Google Slides Plugin. As the title suggests, this is a demonstration of auditing presentation slides with the Google Slides plugin through the Antigravity Plugin Marketplace. Seroter calls it an excellent use case, and his takeaway is that agents and plugins can help keep content up to date.
Next, Announcing Project ProtoStar: Making Publicly Funded Science AI-Ready. Seroter says this is very cool, and he points out that the US invests a ton in research but the output isn't always usable by everyone. This initiative looks like it will help. The post comes from Hugging Face and a set of partners including Carnegie Mellon University, Purdue, Georgia Tech, Arizona State, and George Washington University. The core problem is that publicly funded scientific datasets, models, and code are often scattered across incompatible systems, stored in static or legacy formats, or released without the documentation and metadata needed to reuse them. Project ProtoStar aims to close that gap by encouraging systematic sharing of AI-ready scientific artifacts on open infrastructure, and by running open challenges to demonstrate what becomes possible when those resources are reused. The first challenge, hosted by Carnegie Mellon, runs November 2 through 13 and asks participants to predict climate indicators for 210 US cities using historical weather data through 2018. Teams forecast average annual temperature, total precipitation, the hottest day of the year, and cooling degree days, then compare those predictions against observed weather and simple statistical baselines. They also propose a climate adaptation policy idea informed by what they learned from the data. The broader goal is to build practical evidence about what makes publicly funded science more accessible and useful, and Seroter's reaction is that this is a meaningful step in that direction.
Then comes the InfoQ AI, ML and Data Engineering Trends Report – 2026. Seroter notes that the innovator and early adopter side of this report is stuffed with areas to investigate and apply. The report's key takeaways capture a shift from large language models toward smaller and reasoning models, the movement of AI agents from personal assistants to team and organization-wide systems, and growing interest in system one models like TypeSafe AI's Jev because they offer deterministic and faster inference at lower cost. AI inference is called the year's defining theme, with companies shifting from training models to running inference in production. Evaluations are also getting significant attention as a way to ensure agents do the right thing for the right tasks. In the innovators category, the report highlights agent-to-agent protocols, inference engines such as vLLM, SGLang, and llm-d, harness engineering, spec-driven development and AI software factories, and physical or embodied AI. Physical AI carries a different risk profile because mistakes have real-world consequences, and the report notes that reliability, simulation, safety, and verification become more important when intelligence moves into robots, vehicles, and industrial systems. The early adopter section emphasizes agent orchestration and the shift from interacting with AI to delegating work to it. The panelists describe agents moving from answering questions to planning, using tools, executing actions, and continuing work until a task is complete. For Seroter, the report is a map of where to look next.
Hill-Climbing Agent Quality with Agents CLI Evaluation is the next piece. Seroter calls it a really good post that explains hill-climbing with an agent and what we've learned about each step. The article is a hands-on walkthrough of the Agents CLI evaluation loop. It argues that building an agent that works in a demo is different from building one that stays working across model upgrades, tool changes, and new conversation shapes. The core loop is to write test cases, run them, grade them, read what failed, change the agent, and run again. The post walks through a small weather agent, starting with a baseline mean score of 4.0 out of 5 across four cases, then introduces a realistic regression by accidentally pasting a chef persona as the system prompt. The score drops to 2.75, and per-case analysis shows the chef prompt caused the agent to ignore its weather tool and hallucinate conditions. The article also demonstrates swapping to a cheaper model, which produces the same overall score drop but for very different reasons: the cheaper model holds up on most cases and only fails badly on an off-topic question about a recent event. The lesson is that summary metrics can mislead, and per-case diffs are where the hill-climb actually happens. The post then shows how to schedule evals, alert on score drops, and run automated prompt optimization. It closes with patterns the team has learned, including leading with a small representative dataset, snapshotting full result files rather than scores, separating the dataset from its expectations, treating summary metrics with suspicion, and remembering that more evals do not automatically lead to more fixes. Seroter's note that this explains hill-climbing step by step is well earned.
Next is the New Stack piece, “I don’t think the world needs a new GitHub”: What GitHub’s leaders really think about their new rivals. The title frames a conversation with GitHub's leaders about a wave of new rivals building agentic infrastructure. Seroter agrees with the quoted line: the world doesn't need a new GitHub. But he adds that GitHub has to evolve, and they know it. He says he's rooting for them.
Patrick John Kelly's Prefix Caching and the Hidden Cost of Editing Chat History is next. Seroter says this was new information for him, and the key takeaway is to keep the start of your prompt static. The article explains that each turn in a chat sends the entire conversation history, but model providers can cache the unchanged prefix of that context and charge only about 10 percent of the normal input cost on the cached portion. That savings is fragile. Because the cache only works up to the point where the prompt changes, putting dynamic system instructions or changing values near the top can invalidate the savings for everything after them. The golden rule is to make modifications as close to the end of the conversation history as possible. The article also shows that context compaction has a hidden cost: since prefix caching already saves 90 percent on the static part, rewriting history has to shrink the context by more than 90 percent just to break even on the next turn. It recommends pruning dead file reads or bulky tool outputs right at the end of the latest turn rather than rewriting the entire conversation. There is research into middle-edit recovery and suffix caching on open-weight inference engines, but until model APIs expose those controls directly, keeping the prefix static is the simplest way to stay fast and cheap.
Innovate on your own business terms: Google Cloud spend caps now generally available is the next item. Seroter says he missed these features when reading the release notes a few weeks ago. Now you can set spending caps via API, view a real-time dashboard, and count on automatic resets each month. The post announces that spend caps on Google Cloud Budgets are generally available, with three major updates driven by preview feedback. The first is automation through a spend caps API, accessible from the existing Cloud Billing Budgets API, so organizations can programmatically apply, update, or delete caps and instantly attach them to new sandbox projects. The second is real-time visibility with live cost reports, so when a cap triggers, teams can see the estimated costs that triggered enforcement rather than waiting for standard billing pipelines. The third is automated monthly resets, meaning enforced caps reset and blocked services resume at the start of the next month, while a manual override remains for teams that need a deliberate review first. The post also offers operational guidance: start with strict caps on non-production environments, configure alerts at 50, 80, and 100 percent of the cap, and understand that provisioned throughput is never halted even when a cap triggers, while pay-as-you-go and CUD-covered requests stop. Seroter's framing is that this closes a gap in cost-control tooling.
How I Cut API Response Times by 60% with Go Caching Strategies is next. Seroter's take is that this kind of article feels like knowledge, and it's the sort of familiarity you should have before vaguely asking an AI agent to implement caching. The article describes a Go API whose p99 latency crept above 800 milliseconds because every request hit PostgreSQL for reference data that changed only once a day. Profiling showed about 70 percent of request time was in database roundtrips. The author layers three caching strategies. First, an in-memory TTL cache with a read-write mutex, a janitor goroutine, and generics for type-safe entries. Second, singleflight from the official extended sync package to prevent cache stampedes by allowing only one in-flight call per key when a cache entry expires. Third, HTTP ETags so clients can receive a 304 Not Modified response instead of a full payload when nothing changed. After deployment, the dashboard endpoint dropped from 820 to 310 milliseconds, and the catalog endpoint from 340 to 95 milliseconds. A permissions endpoint stayed uncached because it varied by user and changed on writes, which was the correct call. The post also notes that caching at the HTTP middleware level failed for authenticated responses because per-user cache keys could not be constructed cleanly. The takeaway is to cache close to the data, not close to the transport, and to cache only what is actually stable.
Finally, Why are good developers so wary of agentic development? by Rachel Laycock appears on Martin Fowler's site. Seroter notes that going all-in on a technology or paradigm is a big deal, especially when you're already experienced in that field. Laycock's TLDR is that good developers' caution comes from knowing that built is not the same as working in production. She describes why experienced developers are not hesitant to use coding assistants, but are hesitant to go all-in and let agents do all the execution. A Thoughtworks experiment brought senior engineers, infrastructure consultants, and product people together for a week where no person wrote code and everything was built by agents. The team spent about two-thirds of its time at the whiteboard working out the domain, drawing boundaries, and making plans, and one-third building the harness of tests, contracts, architecture checks, and delivery pipelines around the agents. The application ran, but the build was frequently red, which Laycock says is exactly the uncertainty continuous integration was supposed to remove. DORA's 2025 research found AI adoption is linked to higher delivery throughput but lower delivery stability, so the tension is not unique to one experiment. Laycock argues that the best developers are not nervous beginners on a nursery slope, but people looking at unmarked terrain and asking sensible questions about the route. Leaders need to listen carefully and create room for experimentation, while developers need to bring their judgment and test which of their assumptions still hold. Seroter's comment frames the same idea from the practitioner side: all-in adoption is a serious decision.
Across the list, a common thread is the need for judgment. Evaluating agents before trusting them, understanding caching and cost primitives underneath an AI suggestion, and treating production behavior as the real test all appear again and again. That closes issue 900.
- Daily Reading List – October 30, 2026 (#900)
- Antigravity Plugin Marketplace in Action: Auditing My Talk Slides with the Google Slides Plugin
- Announcing Project ProtoStar: Making Publicly Funded Science AI-Ready
- InfoQ AI, ML and Data Engineering Trends Report – 2026
- Hill-Climbing Agent Quality with Agents CLI Evaluation
- “I don’t think the world needs a new GitHub”: What GitHub’s leaders really think about their new rivals
- Prefix Caching and the Hidden Cost of Editing Chat History
- Innovate on your own business terms: Google Cloud spend caps now generally available
- How I Cut API Response Times by 60% with Go Caching Strategies
- Why are good developers so wary of agentic development?

