↵ select ↓ ↑ navigate esc close

Seroter's Daily Reading: Daily Reading List – October 26, 2026 (#896)

Seroter's Daily Reading ·

Listen: https://blossom.buildtall.systems/663ed2b3572ed337768d81cbc4e498befd53528cf85a60cb7b9541687a5b3696.mp3

Source: https://seroter.com/2026/10/26/daily-reading-list-october-26-2026-896/


Seroter's Daily Reading for October 26, 2026, edition 896, covers Richard Seroter's Daily Reading List. He opens by saying today's list has a ton of interesting advice, especially on team and individual workflows, and the links bear that out: they move from personal AI coding practices to enterprise agent design, and from hard financial questions to a reminder of technology's human stakes.

The first item is My AI Coding Workflow, v3, a post by Helderberto. Seroter liked this one, and made a distinction worth keeping in mind: he is increasingly interested in workflow and process skills rather than skills that just teach a particular product, because the models and harnesses can handle the product mechanics. The post describes the third version of a personal AI coding flow, where steps that used to be prompts have become skills, and reminders have become checks that fail. Helderberto moved off his previous TracerKit tool because a separate tool had its own release cycle; now every step lives in one agent-skills repository. The flow runs from triage through an interview-style grill, spec and ticket creation, build, review, pull request, and retro. Debugging requires a reproducible failing loop before hypotheses are allowed, and the retro step suggests changes to the environment rather than the code.

Next is Introducing new self-learning context enrichment in Knowledge Catalog, from Google Cloud. This preview and the new Knowledge Studio experience turn context curation into a continuous feedback loop. The system analyzes agent and user interactions, finds missing or outdated definitions, and pulls verified context from documents, repositories, and business systems into governed context sets. Google says this makes AI agents nearly four times more accurate, with up to thirty-three percent lower latency and up to seventy-five percent fewer tokens. Seroter had this in his ready-to-read pile for a few days and said he should have read it sooner; he calls it very cool functionality and plans to go hands-on soon.

Then comes Anthropic's Building agents: When to fold subagents into your main agent loop. Seroter frames the question directly: whither subagents? Should more routing and work fold into the main agent loop? Anthropic argues yes, at least in many enterprise deployments. The post describes failure modes like subagent sprawl, the Frankenprompt, and added latency, and suggests that a single agent with progressively disclosed skills and tools often outperforms a multi-agent design. Subagents still have their place for real context isolation, parallel work with fresh windows, or when one loop suffers from context bloat. Anthropic also offers a migration kit and a stage-by-stage approach: build task-completion evals, merge tightly coupled specialists, call remaining subagents as tools, and gradually fold others into skills and tools. Seroter says Anthropic makes a good case.

The next piece is How Google Beam connects military families globally. Google and the USO are bringing HP Dimension with Google Beam to twenty USO locations so deployed service members can have lifelike, 3D video calls with loved ones. The technology captures body language and eye contact, and a new docuseries follows four families from the 2nd Battalion, 37th Armored Regiment during a deployment in Poland. It shows a promotion celebration, a child reading stories, a father seeing his newborn's growth, and a parent witnessing first steps. Seroter says this hit him hard: he lives in a military town and sees how tough it is for families to be apart.

Apollo economist Torsten Slok brings a less heartwarming data point in AI delivers ‘zero financial payoff’ to most companies, Apollo economist says from TechCrunch. Slok looks at the S&P 493, meaning the S&P 500 minus the seven big tech names, and finds flat profit margins and flat revenue per employee, with no evidence AI is improving productivity. He argues that cheap intelligence does not automatically become revenue growth, because business intelligence is mostly judgment. Accountability, trust, culture, unwritten rules, and fairness all shape reputational risk and profits. Seroter says that does not seem like good news, and adds that from what he has observed, few companies have made the structural changes needed to take advantage of AI yet.

That structural point carries directly into the next item, 3 Paths to Delivering Shareholder Returns with AI from Harvard Business Review. The authors study more than six hundred public companies and find roughly three-quarters have invested in AI and moved beyond pilots, but still have not earned a shareholder return. They ask whether companies are simply early or doing something wrong. Seroter's note is succinct: speaking of making structural and capability changes, read this; it is solid advice.

Builder.io asks Why are we still building apps just for humans? Seroter's answer is blunt: you shouldn't be. The argument, drawn from the title and his commentary, is that application logic should be taken out of the user interface so that AI agents can use it, not just people. If a system's capabilities live only behind screens and buttons, agents will not be able to access them. Seroter's take is to get that logic out of the interface where agents will not use it.

Google Workspace has a rollout that Seroter rates ten out of ten, no notes: Preview, edit, and collaborate on markdown (.md) files natively in Google Docs and Drive. The announcement from Google developers says markdown files can now be opened, edited, and collaborated on directly in Google Docs without converting them. Drive gets rendered previews with support for tables and hyperlinks, plus a quick toggle to raw text. Version history, comments, and suggestions work across both human and AI agent contributors. Seroter's summary: you can do full lifecycle work with markdown files without leaving Google Docs.

Next is Patrick John Kelly's Your AI Agent Doesn’t Need to Keep Every File It Reads. The core insight is that many agent harnesses keep everything an agent reads in context until compaction runs, and that bloat raises cost, slows responses, and causes context rot. Kelly analyzed coding sessions and found that between forty and eighty-nine percent of the files an agent reads during a task are never actually used or edited. His proposed fix is post-turn context pruning: keep the material that was legitimately used, prune the rest, and optionally keep only the verbatim line segments that mattered. The subtlety is prefix caching; pruning needs to happen at the tail, right after an exploratory turn, so it does not invalidate the cached prefix. Seroter calls this smart thinking, noting that we are reading a lot of context that is not needed once a turn is done, and all that bloat costs something.

InfoWorld's Stop building agents that can’t change models takes on a related but broader portability problem. The article argues that provider churn is constant, with model retirements and API shutdowns, and that teams often weld their agents to one vendor through SDK types, hosted state, prompt quirks, fine-tuned models, and missing evals. The answer is a stable contract between agents and models, tools and state outside any provider, and evaluations that make a model change a measurable decision. The Model Context Protocol and open formats help keep tools portable, and a real evaluation suite turns a retirement notice into a routine migration. Seroter puts the challenge personally: could you swap models and harnesses in forty-eight hours? You might need to someday, so think about what needs to be in place to avoid starting from zero.

The following item is How Gemini 4 Argon Upleveled Agents on Antigravity: Trillion-Edge Graphs and Custom TPU Kernels. This post from the Antigravity and Google Research team shows what happens when agentic workflows tackle deep systems engineering. Human researchers stayed in control of architecture and acceptance criteria while Teamwork's subagents implemented, tested, and verified. In one sprint they built FarArray, a hardware-accelerated partitioned global address space over RDMA that scales lock-free graph algorithms to trillion-edge graphs, with sixty to ninety times higher throughput in one benchmark and dramatically lower memory use. They also created custom TPU kernels for sparse graph neural networks that cut high-bandwidth memory by up to one hundred eighty-five times and accelerated passes by eight to fifteen times. Seroter's reaction: a year of work done in five hours, and a smart solution to boot? Not bad.

Finally, Thoughtworks contributor Himanshu Sharma writes Don’t ask AI to write the user story yet. The problem is confident gaps: when an AI turns a vague request into polished requirements, it quietly invents business rules that nobody validated. Sharma suggests asking the assistant to sort inputs into three piles: confirmed, assumed, and unknown, and to cite the source for every confirmed item. Missing decisions should be prioritized by asking whether the answer could change scope, change access, or change what the outcome means. The team then brings those questions into refinement, decides them with stakeholders, and writes the story last from a decision record. Seroter agrees with the direction: get questions from your LLM, your team, and yourself, rather than having AI write out requirements based on vague input.

Taken together, today's reading list is a coherent set of workflow lessons. Seroter highlights skills over product recall, simpler main-loop agents with careful context hygiene, portable model architecture, and the human judgment still required before AI writes requirements. He also leaves room for a reminder that the best technology work is not always about efficiency; sometimes it is about letting a deployed parent witness a first step.

  1. Daily Reading List – October 26, 2026 (#896)
  2. My AI Coding Workflow, v3
  3. Introducing new self-learning context enrichment in Knowledge Catalog
  4. Building agents: When to fold subagents into your main agent loop
  5. How Google Beam connects military families globally
  6. AI delivers ‘zero financial payoff’ to most companies, Apollo economist says
  7. 3 Paths to Delivering Shareholder Returns with AI
  8. Why are we still building apps just for humans?
  9. Preview, edit, and collaborate on markdown (.md) files natively in Google Docs and Drive
  10. Your AI Agent Doesn’t Need to Keep Every File It Reads
  11. Stop building agents that can’t change models
  12. How Gemini 4 Argon Upleveled Agents on Antigravity: Trillion-Edge Graphs and Custom TPU Kernels
  13. Don’t ask AI to write the user story yet