Seroter's Daily Reading — #739 (March 11, 2026)

Welcome to Seroter's Daily Reading, an audio summary of Richard Seroter's daily reading lists. This is number 739, from March 11th, 2026. Twelve articles today, and the mix is interesting: a big acquisition, some genuinely thoughtful pieces on agentic engineering and developer workflows, a retrospective on AlphaGo's tenth anniversary, and a concerning picture of tech worker morale.

1. Welcoming Wiz to Google Cloud

Google completed its acquisition of Wiz — the largest acquisition in Google Cloud's history. Wiz was already deeply integrated with AWS, Azure, and Google Cloud, and Google says they'll continue to support multi-cloud. Thomas Kurian frames this as moving cloud security from reactive to proactive. Whether the multi-cloud commitment holds long-term is the real question to watch.

2. Karpathy's Autoresearch

A deceptively simple 630-line Python script that automates the scientific method for machine learning. An AI agent reads its own training code, forms a hypothesis, modifies the code, runs the experiment, evaluates the result, and keeps or discards the change. In one overnight run, 126 experiments. After two days, ~700 autonomous changes and roughly 20 improvements that transferred to larger models. The community response was wild: 35 agents on a peer-to-peer network ran 333 experiments unsupervised, and in 17 hours independently rediscovered ML milestones that took human researchers nearly eight years.

3. The 8 Levels of Agentic Engineering

A maturity model for how developers work with AI coding tools:

Tab complete — Copilot autocomplete
Agent IDEs — Cursor, multi-file edits
Context engineering — deliberate token management
Compounding engineering — codify lessons so every future session improves
MCP and skills — agents access databases, APIs, CI pipelines
Background agents — parallel agents working while you sleep
Orchestration — coordinating agents at the project level
Full autonomy with human oversight

Key insight: your output depends on your teammates' levels too. If your code reviewer is at level 2, their bottleneck becomes yours.

4. Plan Mode in Gemini CLI

Google added a read-only planning mode to Gemini CLI. The agent can navigate your codebase, search for patterns, and read documentation, but cannot modify files. Separates thinking from doing. Includes a new ask-user tool for clarifying questions and supports read-only MCP tools for pulling context from GitHub, databases, or docs during planning.

A detailed architecture for solving GPU underutilization. Combines GKE Autopilot for infrastructure, GPU time-sharing (up to 5 containers sharing one GPU), and vCluster for isolated virtual Kubernetes clusters on shared hardware. Example: a legal research team and a customer support team each get their own Ollama instance serving different models, but sharing the same L4 GPU underneath.

6. When Developer Workflow Discipline Isn't Enough

Individual developer discipline around AI model selection doesn't scale to enterprise. Hundreds of developers making locally rational decisions produce collectively irrational costs. Financial services firms are already pulling AI orchestration into platform engineering — treating reasoning model routing the same way they treat Kubernetes scheduling. The developer consumes the capability; the platform team owns the control plane.

7. 10 Hacks Every NotebookLM User Should Know

Most interesting tips: NotebookLM can now find its own sources via web search. You can join Audio Overview podcasts as a "guest", interrupting the AI hosts with questions. You can configure each chat with its own personality. And you can upload existing slide decks as style templates for new presentations.

8. 10 Years of AlphaGo

Demis Hassabis reflects on the anniversary. Traces how AlphaGo's architecture led directly to AlphaFold (protein folding), AlphaProof (mathematical theorems), and AlphaEvolve (algorithm discovery). 200 million people watched AlphaGo play Lee Sae Dol. Move 37 — a play so unconventional commentators thought it was a mistake — demonstrated AI could find strategies humans hadn't considered. Lee Sae Dol's reflection: "a definitive preview of the AI era, proving it wasn't some distant future, but a reality arriving on our doorstep."

9. How to Quash Your Fear of Messing Up

FOMU — the opposite of FOMO. Instead of spurring action, it creates paralysis. Three archetypes: perfectionists fear reputational damage; people pleasers fear relationship damage; patient performers fear breaking perceived rules. Practical advice: recognize that not acting has its own risk. The "oatmeal vs oatmeat" typo story is a great example of owning a mistake and turning it into process improvement.

10. Secure Error Handling in Go

Core principle: "split brain" error handling. Create a custom error type that separates what the system sees from what the user sees. Your SafeError struct has an Internal field for debugging and a UserMsg field for clients, and Error() returns the safe message. References a real Kubernetes CVE where service account tokens leaked through error marshalling. Practical, security-focused, specific to Go's error-as-values paradigm.

11. Layoffs, Cost-Cutting Shatters IT Worker Confidence

Glassdoor data: positive sentiment among tech workers suffered the biggest year-over-year drop across all industries. Fewer than half gave a positive outlook. 33,000 tech layoffs in 2026 so far, up 51% from last year. IT unemployment at 3.8%. The chief economist: "Tech workers can see the soft job market out there. They know they have less leverage right now."

12. Bring Your Database Tools to the Agent Skill Ecosystem

Google's MCP Toolbox for Databases now has a skills-generate command that converts existing database tools into Agent Skills following the open specification. Define queries in YAML, run one command, get a standards-compliant skill package. Works with Gemini CLI, Claude Code, or any framework supporting the Agent Skill spec. The bridge from MCP to Agent Skills feels important — two complementary standards converging.

The Karpathy autoresearch piece is the standout. The idea that a simple script can rediscover eight years of ML milestones in 17 hours is the kind of result that makes you reconsider timelines. Pair that with the agentic engineering levels framework and the platform engineering piece, and today's list paints a picture of AI development maturing from individual productivity hacks into systematic, infrastructure-level concerns. Meanwhile, the Glassdoor data is a reminder that the humans in the system are feeling the strain.