Third in the How We Build series. Agentic review made code review fast, so the test suite became the gate on shipping. Pulling the first vertical slice out of Pepper took CI from thirteen minutes to two, and gave the agent a repository small enough to hold in its head.

My reading was scattered across Readwise, Karakeep, GitHub stars, and YouTube likes, none of it near my notes. I built a small daemon that routes the missing sources into Readwise, let the Obsidian plugin bring it into my vault, and let a scheduled Claude skill file it. Capture is effortless, organizing is free.

Second in a series on how we build at Vycari, this time on our issue-to-pull-request pipeline. It triages, plans, and builds on its own overnight, but it never merges. A design decision that turned out to mirror the trust principle behind the product itself, and one that quietly covers for the pipeline in public repositories too.

Two years of Letters from Silicon Valley. Year one was written from inside a company I had been at since 2005. Year two ended with me outside it, and the post that turned out to be the hinge was one I thought was about GitHub statistics.

The industry prices and markets AI on the frontier tier, but almost nothing a consumer agent does needs it. I ran the study on my own product. The fabrication everyone blames on small models tracked how clearly we signalled missing data rather than model size, and the big models fabricated more convincingly. Here's what the numbers say.

Issue trackers were built to hold work you couldn't do right now, and to break big work into human-sized pieces. Agents removed both constraints. First in a series on how we build at Vycari, on why three quarters of our backlog turned out to be decisions rather than work.

ChatGPT's launch made the chat window the default AI interface because it matched what those models were good at. Models and use cases have both grown since then, but the box stayed. I argue that chat is a hobbyist's interface, and that agents belong in familiar, purpose built interfaces instead.

Leaving Google after twenty one years, I could recall about a hundred and fifty colleagues, which is roughly Dunbar's number. So an agent read twenty one years of my sent mail and calendar, rebuilt the real list, and several thousand people got a goodbye. What came back surprised me.

Getting real value from an agent today means curating skills, picking models, managing keys and configuring sandboxes. That is not a tool, it is a hobby. There is nothing wrong with hobbies, but most people do not want one. Closing that gap is why I left Google to co-found Vycari.

Today is my last day as a Google employee. It is a Monday, I am at my own desk, and there is nothing on the calendar. The actual leaving happened a while ago. I wrote the note to my colleagues in about twenty minutes, which surprised me. I had been drafting it in my head for weeks, and when I finally sat down to type it, the thing that came out was much shorter than any of the versions I had rehearsed.

This week's reading list explores the shift toward personal, bespoke software and the skills needed to manage AI agents. We also look at the practicalities of prompt injection defense, a shifting four-year timeline of model capabilities, and ambient hardware interfaces like smart rings that are now shrinking our physical inputs.

California's AB2047 would force every 3D printer sold in the state to run state-certified software that detects and refuses to print firearm parts. As an engineer who owns and runs these machines, I explain why that detection software cannot work, why it would have nowhere to run, who actually pays for it, and what the pattern should teach us.

When a mentee told me she wanted to change her role, her team, and her country all at once, I recognized the energy immediately. But hope is not a strategy. The best career plans break big ambitions into a sequence of moves, each one valuable on its own.

This week's reading spans from stunning dark sky photography and Vaonis smart telescopes to immediate practical utility in Postgres database sandboxing and Flipper's new networking gadget. Plus, reflections on the backlash to enterprise tokenmaxxing, Ferrari's custom OLED displays, and the reality of AI-driven layoffs.

Discover how to effectively measure the improvements in your Obsidian agent with a reliable scoreboard for consistent performance evaluation.

Discover the Antigravity SDK: a powerful tool for building independent agent software. Explore its features and get started today!

Discover the transformative features of Gemini Scribe 4.8.0, enhancing automation, task scheduling, and agent autonomy for efficient workflows.

Discover how embracing constraints can spark creativity and innovation in software design, turning limitations into powerful opportunities.

Discover Angine de Poitrine's unique sound that defies categorization, blending microtonal music with a bold visual identity. Uncover the brilliance!

Discover how the Starlink Mini transformed our road trip with reliable internet, enhancing navigation, music, and family conversations on the go.

Explore the Bundled Skills feature in Gemini Scribe, enhancing your Obsidian experience with built-in agent capabilities for effective management.

Discover how to use Gemini Scribe projects to enhance AI collaboration in your Obsidian vault, reducing distractions and improving focus.

Discover the evolution of software development in the age of AI agents. Explore insights, surprises, and the future of building with agents.

Discover the evolution of Gemini Scribe, transforming from a chat plugin to a powerful AI platform with advanced memory and search capabilities.

Explore the debate on AI in writing and what it means for authorship, creativity, and the evolving landscape of literature.

Explore the challenges of autonomous AI in the final part of _The Agentic Shift_, delving into responsibility, safety, and future implications.

Explore how AI agents are transforming industries, from customer support to software development, with real-world case studies and challenges.

Explore the debate between CLI tools and MCP, uncovering the strengths and weaknesses of each for modern AI agents.

Explore the critical role of observability in AI agents, revealing insights on debugging and enhancing agent transparency for effective production.

Explore the inevitable shift from scripts to AI agents in coding. Discover insights on automation, tool access, and setting effective guardrails.

Discover how the 'Internet of Agents' is transforming isolated AI systems into collaborative intelligence through new protocols.

Streamline your research workflow with AI! Discover how to integrate Gemini's deep research capabilities into your terminal for seamless productivity.

Explore the rise of Personal Software: a new era of bespoke tools designed for individual needs, transforming how we solve problems.

Explore the Gemini CLI Policy Engine: a smart solution for balancing AI autonomy and safety, preventing costly mistakes while enhancing productivity.

Discover how I transformed my podcast archive into a powerful AI system with Gemini's File Search, simplifying my engineering process.

Explore the complexities of AI agent frameworks, from state persistence to security, and learn how to choose the right one for your needs.

Explore the significance of an AI agent's attention span and learn to manage its context for optimal performance and reliability.

Explore the new security landscape for autonomous AI agents and learn how to build powerful yet safe and trustworthy AI systems.

Discover how to guide AI agents from prompting to programming for effective decision-making, enhancing their reliability and performance.

Explore how AI agents can remember context, improve communication, and bridge the gap between simple tools and intelligent partners.