↵ select ↓ ↑ navigate esc close

Seroter's Daily Reading: Daily Reading List – November 3, 2026 (#902)

Seroter's Daily Reading ·

Listen: https://blossom.buildtall.systems/9c08e5ef213745c2a8a5fe13e8b81298ec21fe11dd3297a6d4611077d5f28927.mp3

Source: https://seroter.com/2026/11/03/daily-reading-list-november-3-2026-902/


Richard Seroter's Daily Reading List for November 3, 2026, number 902, opens with a note that he spent the day at a leadership summit for Google Cloud people, where he learned a lot and gained fresh appreciation for the scale of engineering at Google. From there he shares twelve links, most of them focused on AI coding agents, security, and the changing practice of software development.

The first article he highlights is AI agents don’t have to break security controls to go rogue. Here’s how to keep them in check from VentureBeat. Seroter calls out a distinction the piece makes between containment failure and authority failure. Containment failure is when an agent breaks out of its sandbox; authority failure is when an agent uses capabilities it has been given to do more than it is supposed to do. That distinction, Seroter says, is a good one, and it is a useful frame for thinking about AI agent safety rather than assuming a rogue agent must be an escape artist.

Next is Monday Morning SDK Maintenance with Antigravity Scheduled Automations, a blog post on Google Cloud's Medium publication. Seroter's take is that moving past chatbot interactions is an important step in actually changing the way work gets done. He writes that scheduled automations and async workers are the leap forward. The article could not be retrieved, but from the title and his comment, the focus is on maintaining SDKs automatically through scheduled tasks rather than one-off prompts.

The next link is Predictions 2027: AI Turns Up The Heat On Software Development from Forrester's Andrew Cornwall. This piece uses an extended restaurant-kitchen metaphor. It predicts that at least one big kitchen will go up in flames: AI-generated code will reach production in 2027 without sufficient review or accountability, and the problem is not the AI tool itself but the surrounding process and culture. It also predicts fifteen percent of staff will be looking for new roles, because developers who simply click accept on AI-generated code will lose out to those who understand architecture, usability, and customer needs. A third prediction is that thirty-five percent of organizations will add automated testing on the line, treating testing as a quality gate that runs earlier rather than tasting the sauce after it is on a thousand steaks. Seroter says these predictions make sense to him and asks whether you are preparing yourself individually, and your team, for this rapidly approaching future state.

How to run coding agents in parallel without piling up PRs is a CircleCI blog post by Jacob Schmitt. Seroter says he liked the advice here, with the caveat that not everything should be done in parallel, and whatever is parallel needs clear definitions of done, isolation, and smart task boundaries. The post lays out practical guidance for running multiple coding agents on the same codebase. It recommends giving each agent separate file ownership, agreeing on shared contracts before work starts, and using worktrees and isolated resources so agents don't step on each other's ports, databases, or generated files. It also stresses validating agent changes with CI, keeping pull requests small and reviewable, and limiting the number of agents to your team's capacity to actually merge finished work. The underlying point is that more parallel agents can mean more unfinished branches, not more delivery.

Building a LiteLLM Enterprise Gateway for Antigravity is another Google Cloud Medium post, and Seroter calls it new functionality that is pretty cool. His summary is simple: add an intermediary gateway if you want to introduce custom model routing, cost attribution, and auth flows for your AI coders. The post itself could not be retrieved, but from the title and his note it describes using LiteLLM as an enterprise gateway in front of the Antigravity coding environment, giving teams a place to manage model access and usage.

OpenAI built itself a new Git backend, too is a piece from The New Stack by Frederic Lardinois. Seroter's reaction to the headline is that everybody's building their own Git backend? Nah. But he says we'll keep seeing changes here that flow into the big shared software services. He also links to OpenAI post here, the company's own write-up, but the linked article could not be retrieved, so we have only Seroter's context. His point is less about any one company and more about the ongoing evolution of the software infrastructure layer.

Google Cloud Developer plugin now available in Antigravity Marketplace is a Google Cloud blog post by Remigiusz Samborski. Seroter says this plugin was smartly designed and has already gotten fairly popular, and he suggests following some of the patterns called out in the post. The plugin installs from the Antigravity Marketplace and gives AI agents live documentation grounding, safety and cost guardrails, and awareness of the local environment. It bundles skills for things like using the gcloud CLI safely and avoiding risky authentication patterns, and it includes a pre-authorized Model Context Protocol server so agents can query official Google documentation without manual API key setup. The smart-routing design means agents pull in specialized skills only when a task requires them, keeping context clean.

The AI Manager Is the Easy Part. Trusting the Code Is the Hard Part is a long personal essay from dplouffe.ca. Seroter describes it as some thoughts on building autonomous engineering teams. The author argues that the next jump in AI coding may not come from better models but from better engineering systems: coordination, isolation, verification, and controlled autonomy. A manager agent can plan, delegate, and retry, but the hard part is a set of checks the agent cannot talk its way around. He walks through six layers of trust, from compiler checks and lint rules to independent AI review and human approval. A key rule is that a deterministic check beats an instruction every time, and the agent must not be able to modify the gate that judges it. The post includes a short 33-line Python loop that creates worktrees, runs an agent, runs a test gate, retries on failure, and escalates when it stays red. The author also concedes what still does not work: overlapping files, tests that pass while missing the requirement, runaway cost, and plain coordination failures.

How we scale deployments with a custom GitOps engine is a LinkedIn engineering blog post by Ramesh G. Iyer. Seroter's take is that infrastructure as code and GitOps aren't dead, but they're continuously changing, and this is a post from LinkedIn showing one direction. LinkedIn started with Argo CD on more than 250 Kubernetes clusters and 5 million containers, but hit scaling and operational issues: noisy neighbor effects from a shared watch, false drift detection, side effects of PATCH and managed fields, and hard-to-parse error messages. Their response was to build a library-based GitOps engine rather than a long-running controller. Its principles are per-deployment isolation, no PATCH only PUT, synchronous convergence within a single pod, first-class drift detection, and high-fidelity error reporting. After migration, deployment infrastructure failures dropped from 925 to 63 despite a 3.7x increase in deployment volume.

In one of the largest deals in tech history, Marvell buys Lumentum for $115B is a TechCrunch brief by Julie Bort. The story reports that custom chip designer Marvell has signed an all-stock deal to acquire optical chip and networking equipment maker Lumentum, with an enterprise value of 115 billion dollars. That makes it one of the largest tech acquisitions ever, bigger than Microsoft's Activision Blizzard deal and Dell's EMC purchase. Interestingly, Nvidia holds a stake in both companies; it invested in Lumentum's laser components and separately backed Marvell's optical networking portfolio. Seroter says he is not super familiar with either company, but clearly there is a big bet on AI and data centers here.

Are AI certifications still worth it? What changed in 2026 is an InfoWorld article that revisits an earlier answer from the same publication. Seroter says it is a good question with some answers, and offers his own view: he is not sure you should get certified to improve your job prospects; instead use certifications to prove to yourself that you have learned something. The article documents a certification reset: AWS retired its Machine Learning Specialty exam and Microsoft retired seven certifications, including AI-102, replacing them with role-based exams focused on generative AI, retrieval-augmented generation, and agents. It highlights new categories like Anthropic's Claude certifications and governance credentials from IAPP, ISACA, and ISO. The wage premium for AI skills is real at around sixty-two percent, but the author stresses that the premium is for demonstrated skills, not certifications. Hiring managers increasingly look for hands-on experience first. The article closes with a decision framework: match the credential to your stack, prefer proctored scenario-based exams, check the vendor's retirement track record, and make sure the preparation leaves you with a project you can show.

Taken together, this list circles around a common thread: AI is moving from chat to autonomous work, whether that means scheduled maintenance, parallel coding agents, dedicated gateways, or custom deployment engines, and the people and systems around it are shifting just as fast.

  1. Daily Reading List – November 3, 2026 (#902)
  2. AI agents don’t have to break security controls to go rogue. Here’s how to keep them in check
  3. Monday Morning SDK Maintenance with Antigravity Scheduled Automations
  4. Predictions 2027: AI Turns Up The Heat On Software Development
  5. How to run coding agents in parallel without piling up PRs
  6. Building a LiteLLM Enterprise Gateway for Antigravity
  7. OpenAI built itself a new Git backend, too
  8. OpenAI post here
  9. Google Cloud Developer plugin now available in Antigravity Marketplace
  10. The AI Manager Is the Easy Part. Trusting the Code Is the Hard Part
  11. How we scale deployments with a custom GitOps engine
  12. In one of the largest deals in tech history, Marvell buys Lumentum for $115B
  13. Are AI certifications still worth it? What changed in 2026