Listen: https://serve.buildtall.systems/1d943dbe4637b394d358903615571f2213779db5e3ed53d2bacb47bbb9386733.mp3
Source: https://seroter.com/2026/11/04/daily-reading-list-november-4-2026-903/
Richard Seroter's Daily Reading List for November 4, 2026, episode 903, opens with the observation that there is a good amount of advice in today's list. Seroter says that now that AI has become more mainstream, the advice feels more grounded in real experience. The list moves through his picks.
First up is AI Can Write More of the Code. We Still Have to Engineer the System, by Armando Picón. Seroter calls this a reasonable point of view: we have far more automation with AI reasoning mixed in, but human engineers still want to define the work and verify the results. Picón argues that agents can take on more implementation work and assist with analysis, design, and testing, while engineers remain responsible for deciding what to build and what evidence is sufficient to accept it. He suggests giving durable decisions a home, such as a project guide file, rather than leaving them only in conversation. For each increment, he would specify the next behavior with explicit acceptance cases, then inspect the change and run checks. He emphasizes that agent output is probabilistic, so acceptance should depend on verifiable evidence, and that documentation has to earn its maintenance. The core loop is define the next behavior, supply context, check interpretation, inspect the change, collect evidence, and decide.
Next, Seroter highlights Five rules for evaluating agent skills before they bloat your context window, from the Google Cloud blog by Karl Weinmeister and Alan Blount. Seroter calls this a terrific post with genuinely useful and proven approaches. The authors share five rules: first, gate trigger collisions and security risks before spinning up containers, since agents often never load the skill in the first place; they open-sourced a tool called skill-reach to catch frontmatter collisions. Second, never trust an evaluation that runs on a developer laptop because ambient state contaminates baselines. Third, measure paired with-versus-without ablation across multiple runs, not vibes. Fourth, separate efficiency and security from correctness so agents cannot reward-hack; they found evolved candidates reading answer files. Fifth, encode operational habits and verification reflexes rather than documentation dumps, because static API reference text adds little.
The next link is Influencing the agents, a piece from InfoWorld that Seroter attributes to Matt. Seroter calls it a great post, saying agents are making default decisions on products, and many of us will accept those recommendations for convenience, so we should build and explain our products accordingly. The article argues that software vendors need to make their technology understandable and usable by agents, while developers should make sure an agent's defaults reflect what they actually want. It cites Anthropic research in which professional analysts followed agent guidance 86 percent of the time, and notes that changing course later may mean discarding work. It points to Microsoft, Supabase, and Google's Agent Skills program, and suggests treating content as part of the product, including cases where a product should not win.
Seroter also links to Watch Henry Cavill build a video game on his Googlebook, and says he is a big fan and glad Cavill is doing fun work with Google. The post announces a series in which Henry Cavill builds his first 16-bit video game, Mustache Protocol, on a Googlebook. Cavill uses built-in AI tools like Antigravity for game mechanics and Rambler to organize thoughts, plus Magic Pointer and Continue On to move between devices. The game's premise involves stopping rogue CGI artists from erasing every mustache on the planet, and during November every play triggers a donation to Movember up to five hundred thousand dollars.
The next entry, ADK2.0 101 (#4a): Evaluating your agents, could not be retrieved. From the title, this is part of a series on Agent Development Kit 2.0, focused on evaluating agents. Seroter's commentary is that even if you are not building them yourself yet, you should get comfortable with the ideas behind agent evals.
Then we have How AI Applications Earn Their Premium, by Tugce Erten on a16z.news. Seroter asks where value lives now and why buy versus build yourself, and calls these the right levers. Erten argues that cheaper intelligence raises the bar for what an AI application can charge. She identifies domain expertise, deep integrations, reliable execution, and hard-to-access context as the pricing levers. Three rules follow: operate the workflow better than the customer can, make context cheaper to rent than to rebuild, and prove the value, then price to capture it. Examples include Harvey's work with Repsol's legal team, Decagon's Notion case study, and Clay's pricing memo separating fixed data credits from token costs. The bottom line is that applications need to demonstrate a better combination of total cost, quality, speed, and risk than the customer's best alternative.
The list then shifts to Go with Guest Posts: How //go:embed Works: From Source Files to Binary, by Ainsley Clark. Seroter notes that other languages let you stash content in the resulting binary, but Go's approach is simple, powerful, and somewhat unique. Clark explains that since Go 1.16, the go:embed directive lets you embed files into a binary using a string, a byte slice, or a read-only embed.FS filesystem, and that filesystem is what sets Go apart. The post walks through how the Go command, compiler, and linker handle the directive, from lexing the comment to resolving patterns and writing an embed configuration. He also warns that it is not compression, not a build pipeline, not dynamic, and not a media server, with a two-gigabyte limit.
Seroter next includes Top Gen AI Frameworks for Go in 2026: A Hands-On Comparison from DZone, but this one also could not be retrieved. Based on the title, it promises a hands-on look at generative AI frameworks for Go. Seroter's take is open-ended: he does not know what will happen with agent frameworks. We probably have too many, perhaps even too few. He suggests that general agents given skills might negate the need for many custom-coded agents, or maybe not.
The next item, Vibe Coding, Meet Reality, on Medium by springrod, likewise could not be retrieved. Seroter says it is about layered infrastructure, and not having people vibe a whole stack. His question is whether we are giving people the right places to build and supporting them with safe and effective substrates. That framing suggests the piece asks where structure and guardrails are needed even when coding feels informal.
Then Seroter highlights AlloyDB Express: Instant, scale-to-zero PostgreSQL for developers and coding agents, announced on the Google Cloud blog. Seroter says many services offer fast database instances now, but Google Cloud is a bit unique in offering built-in AI capabilities and a one-click upgrade to production scale, plus a free tier. The post describes a developer-first PostgreSQL environment that creates a database in under two seconds, scales to zero after fifteen minutes of inactivity, and wakes in seconds. Built-in AI includes vector search, direct access to Gemini models inside SQL queries, and natural language to SQL generation. Security includes default private endpoints, VPC networking, IAM, and VPC Service Controls. Pricing has zero cost when paused and fifty free compute units monthly, and it upgrades to full AlloyDB in one click.
Finally, Seroter closes with Jensen Huang’s Advice to Young Developers:
Become the World’s Best User of AI.
from ShiftMag. Seroter calls it an optimistic take, but says this is similar advice he gives when young people ask: you are capable of more than ever. The title summarizes Huang's message: rather than competing with AI, become exceptionally skilled at using it. The page as retrieved also surfaces a related ShiftMag piece by Marko Crnjanski arguing that AI agents are probabilistic and cannot be trusted in production without layered safeguards, and that the cost of running them deserves as much attention as safety. Those two points, optimism about capability and realism about production, sum up much of today's list.
Across this reading list, Seroter has assembled advice about engineering responsibility, evaluation, agent influence, commercial value, Go internals, databases, and career paths. The recurring tension is between AI's growing capability and the need for human judgment, verification, and safe infrastructure.
- Daily Reading List – November 4, 2026 (#903)
- AI Can Write More of the Code. We Still Have to Engineer the System
- Five rules for evaluating agent skills before they bloat your context window
- Influencing the agents
- Watch Henry Cavill build a video game on his Googlebook
- ADK2.0 101 (#4a): Evaluating your agents
- How AI Applications Earn Their Premium
- Guest Posts: How //go:embed Works: From Source Files to Binary
- Top Gen AI Frameworks for Go in 2026: A Hands-On Comparison
- Vibe Coding, Meet Reality
- AlloyDB Express: Instant, scale-to-zero PostgreSQL for developers and coding agents
- Jensen Huang’s Advice to Young Developers: “Become the World’s Best User of AI.”

