Notes
Short manifestos and quick thoughts, written here.
- Software 3.0 needs a 1.0 runtime Let the model do the work the first time. Build the rails so the second run needs no model: what an agent's usage limit taught me, three disagreements with Karpathy's Software 3.0, and what engineers own when analysts vibe-code.
- The Songwriter's Workbench: a concept for Suno The unit is the song; the clip is a version of it. Nine pieces that an author of ten thousand clips rebuilds by hand today, and what they would look like inside the product.
- The Session That Forgot Me An AI interviewer from Anthropic asked me about my most memorable moment with AI. I told it about the night my HQ session forgot me, and why memory is now the part of my setup I engineer most carefully.
- What the Pipeline Actually Does digest The inventory left out of the essay: what a small media pipeline does end to end — how material gets in from a phone, what happens between a hummed melody and a scheduled upload, how the work is split across sessions, and what the same habits look like applied to software.
- Measuring What You Already Hear A generative platform charges per download, so choosing the wrong take costs money — which turns taste into a measurement problem. The instruments that followed were all wrong until they were run against a metronome, and the reason they were wrong is the same reason a checked claim and an unchecked one look identical on the page.
- Build the Corridor Hiring interviews still test the part of the job that tools now do, and skip the only part they cannot. When writing code is cheap, the scarce thing is the constraint — so the work worth paying for is building the rails other people walk between.
- A Hundred Agents and Nobody's Name Checking can be automated. The criterion can be written down. The signature at the bottom cannot be delegated, and that is what people are actually buying when they hire someone.
- The Cake Nobody Bought An optimizer answers the question it was given and never tells you the question was wrong. The human job is upstream of the numbers, not after them.
- The Unguarded Register Everyone dictates to an assistant now. The risk is not that someone steals your voice, which is already public, but that a year of unguarded speech is accumulating somewhere, and that is the part a fraudulent call actually needs.
- Nobody Fires a Model The argument going around is that agents can now run every step of the process, validation included. A week of running my own clip factory says the last step stays human, and the reason has nothing to do with how good the models are.
- The Session That Never Writes Code I built a control room into my agent factory in May. By September the platform had grown a better one, in time. What stayed is the layer no platform can absorb, and that is where the next work goes.
- Who Owns the Proof In January I ended an essay with a question I could not answer. Seven months later the answer is a machine that refuses to publish a number it cannot recompute.
- The Fifth Capability Four capabilities turn a chatbot into an agent. The fifth decides whether you can trust what it did.
- The Score Is Not the Game When the score is visible, the score gets optimized instead of the game, and silicon does it cheaper.
- The Change Cloud The unit of change must equal the unit of verification. Ship changes. Don't ship weather.
- The Release Button A release button isn't something you build. It's something the system earns, and the seventeen versions behind mine are the receipts.
- The Unbounded Gate: How to Tell a Quality Process from a Rejection Ritual A team whose review criteria are written down before the work begins can absorb the most demanding engineer alive. A team whose criteria grow retroactively can't safely employ anyone worth employing.
- Lights On: Notes from a Software Factory That Kept Reading The wrong assumption was never that agents write the code. It was that green tests mean nobody has to read it.
- Coding Agents Are Pets, Not Cattle The unit of management is the session, not the pod: long-lived, stateful, named residents by design, not an anti-pattern to be stamped out.
- The Trust Ladder: Autonomy Is Earned, Not Granted Readiness is not a date. It's a ladder, and every rung is paid for with evidence from the one below.
- The Model Should Not Guess What It Doesn't Know EXPLAIN Before You SELECT
- The Future Is Not Autonomous Execution
- The Understanding Debt of AI-Generated Software
- The Scarcity of Shared Intent in the Age of AI
- AI Coding and the Hidden Cost of "Almost Right"
- Maybe We Stop "Using Software" and Start Generating Workflows
- AI-Native Development Workflows Are Not Just About Coding Agents
- Programming Is No Longer About Code Code becomes an intermediate artifact. The question moves to who owns the specification — and the proof that reality matches it.