↵ select ↓ ↑ navigate esc close

Seroter's Daily Reading: Daily Reading List – October 9, 2026 (#885)

Seroter's Daily Reading ·

Listen: https://blossom.buildtall.systems/515dd256562d870b559ac5f344c3cfc9e61a3400b211d21cffcefdcfe0069176.mp3

Source: https://seroter.com/2026/10/09/daily-reading-list-october-9-2026-885/


Welcome to Seroter's Daily Reading for October 9, 2026, edition number 885. Today's list doesn't have any product focus, but rather techniques and topics to think about. Seroter kicks things off with a piece on Building scalable, agent-friendly APIs for AI applications from The New Stack. His take: APIs are just APIs, right? Regardless of client? Sort of, but agent calling patterns should force us to revisit a few assumptions. The article argues that AI agents interact with APIs in ways that break traditional assumptions—long-running loops, tool orchestration, and the need for structured, machine-readable contracts. So designers need to think about streaming, tool definitions, and state management differently.

Next up is a blog post by Charity Majors titled If your team is happy, are you doing a good job? Seroter says there's a lot here that needs to be said, and his own framing is that our job as managers is to get business outcomes, and ideally making it awesome for our teams too. Majors digs into how psychological safety became unmoored from performance, tracing the ZIRP-era management philosophy that equated happiness with success. She argues that learning often feels uncomfortable, and that a manager who only tries to keep people happy may not be driving the business forward. The post is a call to return to the original meaning of psychological safety: enabling learning and high performance, not avoiding all negative feelings.

Geoffrey Huntley's bluntly titled post, if your team is too busy doing their 'normal job' to experiment with AI, you're preparing them to be replaced, gets straight to the point. Seroter says it needs to be said, and adds that if you plan on being employable in tech, you should be proficient with AI tools. Huntley argues that AI usage is no longer optional in software engineering. He recommends managers carve out time for professional development, build agents, and even suggests that hiring managers should reject candidates with a poor attitude toward AI. The post includes a stark hiring rubric and a quote from Bill Gates about job destruction once tasks cross reliability thresholds.

From the world of JavaScript runtimes, Deno is joining Cloudflare. Seroter congratulates the team, but notes this looks like the end of Deno as a standalone runtime, since it will only be supported for the next twelve months before development ends. The hope is that it keeps going via pure open source. The announcement explains that the Deno team will work on Cloudflare Workers and Durable Objects, with Deno Deploy shutting down after six months. The focus shifts to a shared platform built around the celld project, which uses Durable Objects for stateful agent harnesses.

Harvard Business Review has a piece called The 6 Essential Skills Your Workforce Needs to Keep Pace with AI. Seroter says this part seems pretty clear: those who do strategic and systems thinking, mixed with applied expertise and judging AI output, will be best set up for the future. The article, based on Gartner research, argues that upskilling should focus on a small set of skills rather than an ever-expanding checklist. Unfortunately the fetched article text doesn't include the full list, but the summary makes clear it's about combining strategic thinking with applied AI expertise and the judgment to evaluate AI output.

On keeping LLM judges honest, the Research-Driven Engineering Leadership newsletter asks What does it take to keep an LLM judge trustworthy in production? Seroter's question: can the LLM just be the judge instead of requiring human evaluation? His answer: that judge needs curated training data and more, which means humans still play a role. The piece dives into Netflix's experience evaluating recommendation explanations with LLM judges. They found that a judge needs a rationale-labeled benchmark, weekly human review to catch drift, and careful tuning on reasoning rather than just labels. Even with automated monitoring, human reviewers surfaced gaps the metrics missed.

Google's developer blog presents The Outer Loop, Insights First: An Ambient Quality Agent That Diagnoses Your Production Agent. Seroter notes that some are already experiencing it, but pretty soon many will find out that all green dashboards aren't telling you that your AI agent is falling apart. This quality agent helps find those signals that the agent is off track. The tool, called AQuA, runs 24/7, samples production trajectories, grades them against a checklist, clusters failures, and diagnoses root causes against deploy-time source snapshots. It's a way to catch silent failures like an agent bypassing a seat availability check while still returning HTTP 200.

Juan Cruz Martinez shares 7 Things that I Encourage as a Manager. Seroter says there are all sorts of ways to manage, and if you're new at it, consider these situations to encourage. The list includes using one-on-ones for anything going on, protecting time from unnecessary meetings, disagreeing with the manager, giving honest feedback, saying no, asking why we're doing this, and reaching out before burning out. Each is about building trust and improving the team.

Google Open Source Blog asks How much does it cost to use open source software? Seroter calls this unique research, and mentions that Sophia Vargas, who wrote it, is on his team and does great exploration of OSS. The study looked at how long it takes an engineer at Google to update an open source package. The majority take four hours or less, but the long tail can stretch to days or months. Key predictors of update complexity were local patches and customization, the number of package dependencies, and the number of upstream contributors—not the age of the package or how many teams use it.

Stripe's blog has Why I tried to kill token billing (and why we kept it), by Scott Woody of Metronome. Seroter asks: is output-based billing much better than pricing products with passthrough token costs? Woody argues that token billing is useful on the backend for tracking costs but bad as a customer-facing model because it exposes your margins and invites commoditization. He suggests unified credits as a stepping stone and advocates output-based pricing—charging per resolved conversation or generated image—while calling outcome-based pricing mostly a myth except for a few large contracts.

The DX newsletter weighs in on Don't measure AI Code Percentage on its own. Seroter asks how we should think about this metric, or not, and says this seemed like a good take. The authors argue that AI Code Percentage is a useful lens for understanding outcomes, but a bad KPI. Adoption is already saturated at over 95% weekly usage, and the percentage of AI-authored code has risen from 19% to 52% in a year. Gaming the metric is easy, and two teams with the same percentage can use AI very differently—one as autocomplete, another as full delegation. They recommend reviewing AICP quarterly, alongside throughput, quality, and developer experience, and paying attention to autonomy and code survivorship.

Finally, Charlie Guo's Agentic Primitives 101 offers a review of the building blocks for long-running agents. Seroter likes reviewing the foundation because it keeps us from freaking out over the perceived complexity of the higher order thing. The piece breaks primitives into doing the work—instructions and skills, tools and connectors, sandboxes; continuing the work—sessions, compaction, schedules and events; and delegating the work—subagents. It also looks ahead to memory, proactivity, and identity as emerging patterns for ambient AI.

That wraps up Seroter's reading list for today. We saw a strong theme around AI quality—judges, ambient agents, and metrics—plus management advice and open source cost research. Until next time.

  1. Daily Reading List – October 9, 2026 (#885)
  2. Building scalable, agent-friendly APIs for AI applications
  3. If your team is happy, are you doing a good job?
  4. if your team is too busy doing their ‘normal job’ to experiment with AI, you’re preparing them to be replaced
  5. Deno is joining Cloudflare
  6. The 6 Essential Skills Your Workforce Needs to Keep Pace with AI
  7. What does it take to keep an LLM judge trustworthy in production?
  8. The Outer Loop, Insights First: An Ambient Quality Agent That Diagnoses Your Production Agent
  9. 7 Things that I Encourage as a Manager
  10. How much does it cost to use open source software?
  11. Why I tried to kill token billing (and why we kept it)
  12. Don’t measure AI Code Percentage on its own
  13. Agentic Primitives 101