↵ select ↓ ↑ navigate esc close

Seroter's Daily Reading: Daily Reading List – November 5, 2026 (#904)

Seroter's Daily Reading ·

Listen: https://serve.buildtall.systems/f6c7505cb4a79c92350431ca38ca7cc609ed7c293204f92b33d4c6407a28f2c5.mp3

Source: https://seroter.com/2026/11/05/daily-reading-list-november-5-2026-904/


This is Seroter's Daily Reading for November 5, 2026. Richard Seroter opens by noting that the AI SDLC has been a hot topic this week among his customers. Many people want blueprints, but he isn't sure many general purpose ones exist. Work to do.

First up is How two engineer side projects became the agent taking action inside Dropbox. Dropbox engineers Daniel Reiter Horn and Rich Chan had each built side projects to let a coding agent work directly with their Dropbox files. Daniel's sandbox gave an agent a place to run software for 3D modeling, while Rich's desktop app prototype added a chat interface and permission controls for turning receipts into expense reports. The merged system brings those ideas into the new Dropbox experience. Agents work in an isolated workspace, access files through Dropbox's existing permission system, and every change is tied to a conversation so people can review or undo it precisely. Permission checks live outside the sandbox, and the agent never grants itself broader access. Seroter calls this thoughtful work, both in how they approached user interactions and in evaluating permission handling.

Next, How Context Drift Impacts Your AI Agents. Patrick John Kelly ran an experiment to separate context rot from context drift. Context rot is when a session gets too long and noisy, but context drift is when underlying files like instructions, skills, docs, and code fall out of sync with each other. He found that injecting fifty thousand tokens of unrelated chat history caused zero rule violations, but conflicting instructions were a different story. In seventy-three percent of runs, the model noticed two sources disagreed in its internal thinking, yet it mentioned that conflict to the user zero percent of the time. Instead it silently picked a winner, often a task-specific skill file over a root instruction file. When Kelly gave the model a small read-only tool to log inconsistencies, it logged nearly ninety-nine percent of conflicts with no extra instructions and no real token cost. Seroter says this is a case of moving from what if to getting hard data, and that LLMs swallow context conflicts unless you force them to shine a light there.

The Hidden Productivity Trap in an Engineer’s Workday could not be retrieved when this episode was prepared, so we only have Seroter's framing. He notes that interruptions are a fact of life, but asks whether you are giving yourself or your team windows to just stay heads down. Based on the title, the trap likely involves how interruptions and context switching quietly undermine an engineer's productive output, though the article's own argument isn't available to us.

The November 2026 ThoughtWorks Radar is volume thirty-five, and Seroter calls it probably his favorite tech assessment, saying each release surfaces a handful of techniques or products he has never heard of. The Radar organizes blips into adopt, trial, assess, and caution rings. This edition's themes include harness engineering at scale, where the practices and tools around AI agents mature into a more cohesive discipline. Another theme is beyond frontier models, with open-weight models and local coding models becoming credible alternatives for many tasks. A third theme covers putting more agents to work, including parallel coding sessions and longer unattended agentic builds. The fourth theme is constructing the AI factory, where Kubernetes and related infrastructure become the economic foundation for running AI workloads at scale.

Powering open infrastructure for autonomous agents is a Google Cloud announcement that three projects started at Google have been accepted into the CNCF Sandbox. Agent Substrate, gVisor, and OpenRL are being donated to the Cloud Native Computing Foundation. Agent Substrate decouples stateful agents from compute, snapshotting idle agents to storage and restoring them in under five hundred milliseconds to reach ten times container density. gVisor provides lightweight application kernel isolation for untrusted code, and OpenRL orchestrates reinforcement learning post-training loops on Kubernetes. Agent Substrate has also reached its one point zero milestone and is generally available on GKE. Seroter points to this as more open source components for the emerging agent platform, noting that several key ones were just given to the CNCF.

The next link, titled More here, is a companion Google Open Source Blog post that goes deeper on sandboxing untrusted AI agents at scale. It explains the two big infrastructure bottlenecks: idle resource waste for interactive agents that mostly wait, and control plane overload for training or evaluation fleets. The post describes the actor and worker model in detail, where actors hold state and workers provide reusable compute. It also walks through the lifecycle of an agent turn, from request parking through sub-second restore to suspension on idle, and explains why gVisor and microVMs are used as isolation boundaries. The post frames the donation as a neutral vendor-independent foundation for agentic infrastructure, with secure isolation, agent runtime, and training orchestration as three layers.

Agentic Commerce Protocols — UCP and AP2 could not be retrieved, so we only have Seroter's context. He asks if you are tracking these protocols for commerce and payment, and says this article helped him understand them better. The title suggests it explains two open protocols called UCP and AP2 for agentic commerce, but the specifics of how they work are not available to us.

Eight reasons human developers are better than coding agents is the counterpoint to an article last week about why agents are better than humans. The author stands by the claim that coding agents write code better, faster, and longer, but argues software development is much more than code. Among the eight reasons: only humans can know non-code context like a feature being the CEO's pet project, only humans can be accountable when something goes wrong, only humans can say no to a bad idea or yes to a new one, only humans can say I don't know, only humans can initiate a project, only humans have taste, and only humans can say whether the application is right. Seroter says these points are mostly around agency and accountability, not capability.

Apple is reportedly cutting budgets for its internal AI usage. According to Bloomberg, Apple lowered the daily AI spending cap for engineering teams from three hundred dollars to two hundred dollars in October, with further cuts to one hundred dollars a day and possibly as low as twenty-five dollars outside engineering. The article ties this to the short-lived tokenmaxxing frenzy earlier in the year, when companies pushed staff to use AI as much as possible before the bills came due. Seroter calls the cuts understandable, but says it may not be sustainable if companies expect employees to fully capitalize on AI in their work.

A new Go API for JSON announces that the encoding/json/v2 and encoding/json/jsontext packages are stable and generally available in Go 1.27. The new packages address longstanding behavior flaws in the original encoding/json package, including invalid UTF-8 handling, duplicate object names, nil slice and map serialization, case-insensitive unmarshaling, and inconsistent method calls. The new API separates syntactic encoding from semantic marshaling, and introduces streaming interfaces like MarshalerTo and UnmarshalerFrom. Unmarshal performance can improve up to ten times. The Go team implemented the original package in terms of the new one under the hood so both versions can coexist and migrate gradually. Seroter says Go takes backwards compatibility seriously while still pushing performance forward, and calls this a great blog post.

AI-assisted SDLC in a regulated industry: case study is a detailed look at running a coding agent inside an audited delivery process. The team uses Claude Code to write code, requirements, and tests on a multi-tenant SaaS platform, but the agent has no permission a junior engineer wouldn't have. Three levels of CI gates and five human decision points tie every change to a requirement, a test, and a release. Requirements are treated as code and versioned in the repository, with traceability checked on every pull request. Rules, skills, and hooks also live in the repository and change only through reviewed pull requests. The article emphasizes auditability, quality gates, and telemetry that pairs cost with outcomes. Seroter says he has seen flavors of an AI SDLC personally, close up, and from afar, and mostly it seems about pacing and what's agent-led versus human-focused, with a few variations mixed in.

Gemini Omni 1.2 Flash: Fast, cost-effective, and reliable video generation and editing now generally available is a Google Cloud announcement for a video model aimed at everyday business videos like product clips, virtual try-ons, and ads. The price drops thirty percent to seven cents per second of 720p video, clips can run up to fifteen seconds, and the model follows image and video references more closely. It also offers faster drafts at lower resolution with upscaling, more natural speech and lip-sync, and conversational editing of the same clip instead of starting over. Outputs carry an imperceptible SynthID watermark. Seroter says it looks good, with cheaper, more consistent results, more natural speech, and faster generation.

MCP vs API vs CLI: How Should AI Agents Connect to Your App? is an extended look at the options for making a web app agent-friendly. The author builds all three surfaces on top of the same server logic and argues that MCP deserves most of the care because it standardizes both the server and client sides. MCP handles login, permissions, and instructions, and is the only option that can route a human confirmation directly to the user with an elicitation instead of letting the model answer for them. APIs still win for bulk jobs and scripts, while a CLI is probably unnecessary in most cases. Seroter says the author undersells the CLI and doesn't factor in token consumption, but the overall assessment is solid.

Finally, When Building Gets Easier, Decisions Get Harder could not be retrieved, so we rely on Seroter's context. His take is simply, sure, the product manager job didn't just get easier, and now there are even trickier things to work on. The title suggests the article argues that as building software becomes cheaper and faster, the harder problem shifts to deciding what should be built and what should be killed, but the article's own content is unavailable.

Across the list, several tensions recur. Agents are increasingly capable and open infrastructure for them is maturing, but human agency, accountability, and judgment remain central. Cost and productivity are also recurring questions, whether in Apple's budget cuts or in the token consumption that agent integrations still need to manage.

  1. Daily Reading List – November 5, 2026 (#904)
  2. How two engineer side projects became the agent taking action inside Dropbox
  3. How Context Drift Impacts Your AI Agents
  4. The Hidden Productivity Trap in an Engineer’s Workday
  5. November 2026 ThoughtWorks Radar
  6. Powering open infrastructure for autonomous agents
  7. More here
  8. Agentic Commerce Protocols — UCP and AP2
  9. Eight reasons human developers are better than coding agents
  10. Apple is reportedly cutting budgets for its internal AI usage
  11. A new Go API for JSON
  12. AI-assisted SDLC in a regulated industry: case study
  13. Gemini Omni 1.2 Flash: Fast, cost-effective, and reliable video generation and editing now generally available
  14. MCP vs API vs CLI: How Should AI Agents Connect to Your App?
  15. When Building Gets Easier, Decisions Get Harder