↵ select ↓ ↑ navigate esc close

Seroter's Daily Reading: Daily Reading List – October 22, 2026 (#894)

Seroter's Daily Reading ·

Listen: https://blossom.buildtall.systems/51ca0f25f815de315e95671f7e362b52230ac1c2e4ff878056106418e005b72f.mp3

Source: https://seroter.com/2026/10/22/daily-reading-list-october-22-2026-894/


Seroter's Daily Reading. Today's post is Daily Reading List – October 22, 2026 (#894). Richard Seroter opens with a blunt warning: What a mess of a week around tech security. He counts at least a dozen vulnerabilities or leaks across Linux, source control, runtime, PaaS, even certificate authorities. That security thread returns at the end, but today's list begins with AI infrastructure and the changing shape of engineering work.

The first link is Integrating TorchTPU into TorchTitan and Optimizing Models for TPU. Google and Meta have co-developed a fork of Meta's TorchTitan reference trainer that runs on Google Cloud TPUs. The post details SimpleFSDP, a compiler-friendly sharding layer that avoids FSDP hooks, a Graph Trainer that captures the whole training step as one XLA program, and Pallas kernels such as Splash Attention and fused losses. The results are significant: FLUX reaches about ninety-nine percent of JAX MaxDiffusion throughput, and Qwen three two thirty-five B holds just over thirty-two percent MFU on a thousand-chip slice. Seroter calls it more work between Google and Meta to scale PyTorch on TPUs.

Next, How to Use AI to Become a More Effective Engineering Leader. This adapted book chapter by Gregor Ojstersek argues AI's biggest opportunities for leaders sit outside coding: drafting design docs, reviewing specs for ambiguity, summarizing long documents, preparing for meetings, translating technical language for business audiences, finding internal knowledge, coaching team members, making data-driven decisions, and automating repetitive work. It also cautions against transcribing one-on-ones and stresses that AI should assist judgment, not replace it. Seroter's take: in October twenty twenty-six we shouldn't have to say don't just use AI for coding, but this is a good reminder that it's useful for many other things.

Ethan Mollick's 30 Months on the Exponential (in Charts) marks the release of his new book Co-Existence. Mollick notes that GPT-4 was once the single best model; since then twenty different models have held the lead, and a hundred eighty models now outperform the original GPT-4, with nineteen able to run on a smartphone. The METR Long Horizon Benchmark ran out of headroom after showing AI could do seventeen hours of human work. Prices for graduate-level question answering fell from thirty cents to four hundredths of a cent per correct answer. Mollick adds a new human-AI working style, the conjurer, who summons agents and evaluates results. Seroter says it's been a wild couple of years, and we're just getting started.

The next item is Introducing the Gemini API CLI. The linked page could not be retrieved. Seroter describes this new experimental CLI as a no-dependency binary that your favorite tools can use to interact with the Gemini API. The title suggests a lightweight command-line interface for Gemini, but without the fetched text there are no further details to report.

Shopify's We gave AI agents a mobile dev environment describes Remote Dev, a system for letting the River agent build and test mobile apps without maxing out laptops. Shopify keeps a pool of warm remote Macs and Linux Android emulators; each River session claims its own environment, builds in the background, reproduces bugs, collects screenshots, and creates pull requests. Live Preview streams simulators to a browser so non-mobile developers and designers can try the app. Adoption of River for mobile pull requests grew from six percent to sixty-one percent. Seroter calls it a cool solution that gets processing off the local machine.

Powering the Agent Development Lifecycle with any framework, from ADK to LangChain and beyond covers a new extension system for Google's Agents CLI. Version one point five introduced overrides and new commands without forking the tool, and version one point eight ships a first-party LangChain extension. A scaffolded LangChain project has no ADK dependency yet still deploys to Cloud Run, runs evaluations, and uses the same infrastructure commands. The post also shows how an enterprise can wrap deploy with an SBOM scan. Seroter recommends this helpful CLI for agents built with frameworks other than ADK.

CircleCI's 5 takeaways from the State of Software Delivery Q3 Pulse report examines twenty-two million workflows. Builds are passing more often, finishing faster, and recovering sooner, but median throughput isn't improving. Feature-branch activity grew forty percent while main-branch activity grew only seventeen percent. The top five percent of teams run ten point four times more main-branch workflows than the median, work across fifty-five percent more branches, and land changes in far fewer CI cycles. Seroter's summary: teams are coding more but shipping at the same rate, and the leading performers are getting clear separation from the rest.

Inside Agent Compute Engine: Building the compute foundation for the agentic era is Google Cloud's deep dive into a new service built for agent workloads. It combines VM isolation with container speed by giving every agent a hardware-isolated microVM. Security is built around zero credentials inside the sandbox, restricted networking through a mediated proxy, and four separate identities so the untrusted environment never holds the keys that open the door. Because most agents spend most time waiting, idle detection checkpoints them to near-zero resource cost; active microVMs share deduplicated memory pages. Seroter calls it micro-VMs and more for isolated, fast compute the right way. He links to a companion piece, More here, from The New Stack; the fetched page is mostly site navigation, but its headlines point to coverage of Google turning the infrastructure behind Gemini's agents into a cloud service.

CIOs unprepared for agentic workforce shift reports on a Gartner survey of two thousand two CIOs. Only four percent have plans in place to revamp roles for a human-agent workforce, even though those that do are two point four times more likely to be top AI performers. Only thirteen percent report significant value from AI investments. Gartner's Alicia Mullery warns that without a ready workforce, costs will play out because teams can't effectively deliver. Seroter responds with an oof and asks if it's really only four percent.

Expanding Claude availability and deployment choice on Google Cloud announces a new Anthropic-managed deployment option alongside the existing Google-managed option. Both offer day-zero access to Claude models, the same pricing, and Google Cloud billing. The difference is who acts as data processor: Anthropic under its terms, or Google under customer agreements. Anthropic-managed gives immediate access to Claude platform features like managed agents, skills, and code execution, while Google-managed adds VPC Service Controls, CMEK, and data residency. Seroter calls it a cool option for teams that want day-zero Anthropic features and Anthropic terms with Google Cloud's management layer.

Finally, 1Password launches a way for AI agents to log in on your behalf without seeing your password. The new 1Password for Agents platform lets an agent request a login, proves its digital identity to 1Password, prompts the user for confirmation, and then fills credentials without exposing them. Supported agents include Claude Duet, OpenAI Dots and ChatGPT Work, Grok Bot, Meta's Muse, and Town. TechCrunch notes banks are already detecting agentic activity inside customer accounts and often can't tell if it's authorized. Seroter connects it to a disastrous week for security and says this is one way to limit what your agent has knowledge of.

Across today's list, a clear pattern emerges. The infrastructure for agents is maturing quickly, from compute sandboxes and CLI frameworks to credential protection. But the CircleCI data and Gartner survey both suggest most organizations are starting more AI-enabled work than they can finish or govern. As Seroter's opening warning makes clear, security remains the unresolved undercurrent.

  1. Daily Reading List – October 22, 2026 (#894)
  2. Integrating TorchTPU into TorchTitan and Optimizing Models for TPU
  3. How to Use AI to Become a More Effective Engineering Leader
  4. 30 Months on the Exponential (in Charts)
  5. Introducing the Gemini API CLI
  6. We gave AI agents a mobile dev environment
  7. Powering the Agent Development Lifecycle with any framework, from ADK to LangChain and beyond
  8. 5 takeaways from the State of Software Delivery Q3 Pulse report
  9. Inside Agent Compute Engine: Building the compute foundation for the agentic era
  10. More here
  11. CIOs unprepared for agentic workforce shift
  12. Expanding Claude availability and deployment choice on Google Cloud
  13. 1Password launches a way for AI agents to log in on your behalf without seeing your password