I rebuilt a database table and it silently deleted two columns I needed. No error. Every test for the thing I was changing passed. Here's what actually caught it.
Three bugs that would have shipped clean. None of them threw an error. The only thing that caught them was a test written to fail first, against the real thing.
There's a file in my project whose only job is to stop me from making one mistake. I made it three times in a row. The difference between a rule and a wall.
Thirteen features shipped green. The designs were all wrong. The system made being wrong cheap—but only on one side of the work.
I once mapped the civil war between human interfaces and machine-native ones at the scale of companies. Then I fought both sides of it in a single afternoon, in a single pipeline — and a human had to close the seam.
I went looking for other cooks and found closed kitchens — not hostile, just complete. The Bear turned out to be the show that named what's actually missing in agent-assisted work, and what it might cost me to build the alternative.
I pasted a two-paragraph plan into a terminal and walked away for two days. The strategy package that came back was good. Then I asked Claude to do it again from a template, without me in the feedback loop. The template version landed at 70-80% of the quality in 10-15x less time. The gap — and the loop that closed it — has a name from another century.
I sent Adam Bender's Google I/O talk to Claude with a note calling it the most insightful take I'd heard on AI and our industry, and asked for a LinkedIn caption. The draft came back confessing things about my work that weren't true. The story of why is the same story Bender is telling on a Google I/O stage.
I spent a session trying to map my skills across sixty-five projects and the output came back dressed up as a senior engineer. That wasn't right. The correction I typed back is the one I think a lot of people are currently refusing to type back.
Supabase added a 'Copy prompt' button that bypasses their entire setup guide. It's a small UI change that reveals a big design shift — and it pushed me from 'context engineering' to 'probability engineering' as the real frame for agent-native tooling.
Etsy and Wayfair are already processing agentic checkout through Google's Universal Commerce Protocol. OpenAI and Stripe shipped a competing one. Cart and checkout are becoming infrastructure services — and the standards war is already here.
I pointed an AI agent at a product initiative and didn't write a single line of code or a single word of copy. Two days later: 11 prototype pages, 4 strategic documents, cross-industry research with 30+ citations. And a trust problem I still can't solve.
I've been asking 'can an agent do this?' so reflexively that I stopped noticing when the answer should have been 'but should it?'
I just mapped the civil war inside every agent system. Both paths sound reasonable. But what if splitting investment across both means neither gets finished?
Every agent system is fighting the same battle: learn to navigate human interfaces, or demand native ones. The industry is doing both. The question is who adapts to whom.
I built a Constitutional AI Governance Framework. Thirteen articles. HMAC attestations. Democratic amendment processes. Every validation function returned hardcoded perfection. The governance thinking was real. The code was theater. Here's what survived the extraction.
GitHub wants agents to explore. Google wants websites to hand them a menu. Two competing philosophies for the same problem—and the one you start with will shape your migration cost when they inevitably converge.
I had a 46-citation research paper about autonomous documentation. Academic frameworks rarely survive contact with a real codebase. So I asked an agent to turn theory into working code—and watched what happened.
The problem with AI coding assistants isn't capability—it's coordination. A single agent can write code. But who checks the work? Who remembers what was decided? I built a system where specialists implement and verifiers catch drift.
I wrote a voice guide to help AI match my writing style. It worked too well—the AI learned the example phrases, not the principles behind them. Here's how I fixed it, and what it taught me about the difference between describing a voice and understanding one.
How many ideas die in the space between waking and coffee? Vercel's Agent Skills announcement made me think about what changes when engineering judgment becomes installable—and who gets to build things when the execution gap narrows.
We're not short on people who can chat with a bot. We're starved for people who can deconstruct a business process into atomic units an AI can actually execute. That gap has a name now.
College students aren't just using AI chatbots anymore. They're building automation systems, running local LLMs, and treating software engineering as a just-in-time capability. The 'Chat Terminal' era is over.
Wade Foster doesn't send memos about AI. He runs hackathons and show-and-tells. That distinction matters more than most CEOs realize—and it's the same thing I've been telling my own teams.
Every e-commerce platform is racing to add AI features. But what if the real opportunity isn't AI features—it's AI architecture? What if the store itself could generate in real-time?
If I handed you my camera right now—same lens, same settings, same light—would you get the same shot? The tools are available to everyone. The output isn't.
I used to say 'just one more level' until 3am. Now I say 'just one more version.' The game changed. The compulsion didn't.
It's Thanksgiving. And while I'm grateful for the usual things—family, health, increasingly creative leftover sandwiches—I need to take a moment to thank the team that really made this year possible. They don't eat. They don't sleep. And they never ask for PTO.
Instructions arent enough. To make agentic workflows reliable, I had to build a meta-agent to police my coding agents. Welcome to the unglamorous world of AI Ops.
Agentic software is powerful, but it needs guardrails. Im finding the most important work isn't coding, but architecting the systems that constrain the code.