Building things in the AI era — software, hardware, and projects for the long game, one decade at a time. Field notes from a solo builder shipping real products with AI agents. No hype, no tutorial-farm noise. Just what actually works, explained clearly and tested against the evidence.
Anthropic released something on 21 October that most people will skim past, because it looks like another assistant. It is worth ten minutes, because the thing it changes is not speed. It changes who decides what you spend your morning on. It is called Claude Duet. Before arguing about it, here is what it actually is. What Claude Duet is, in one paragraph Duet is a proactive agent. Instead of waiting for your question, it watches the apps you connect and comes to you: it scans your inbox, calendar, invoices and messages every few hours, writes a short morning brief of what it thinks matters, and messages you during the day when something needs doing. It also works while you are away — it runs on a computer of its own in the cloud, so it keeps drafting replies or chasing unpaid invoices after your laptop is shut. In Anthropic’s own words, it is “designed for judgment — deciding what needs you, what can wait, and when to check first.” ...
An app can ship, work perfectly, and pass every test — and then stop working because somebody else changed a string. Not your code. Not your deploy. A model name in an API call. The endpoint stayed up; the model behind it was retired, and the API politely returned a 404 in a JSON envelope, as though your request had always been malformed. This is not a rare edge case. It is the normal operating rhythm of a market where models are released, superseded and withdrawn on a schedule you do not control. And it lands differently on a solo builder than on a company: a team gets a vendor email that routes to a platform group and a migration ticket. One person gets a user message saying the feature “doesn’t work anymore.” ...
Earlier this month, the CEO of Perplexity mentioned something in passing that deserves more attention than it got. Aravind Srinivas said there’s one solopreneur — literally a one-person company — spending upwards of a million dollars a year on the product, and making about $3 million a year doing it. The story got retold as an AI-leverage anecdote. One person! Seven figures of spend! Automate everything! Which it is. But the retelling skips the only number that matters, because the two mentioned in the same sentence are not the story. The story is their ratio. ...
Most advice about the first hire answers the wrong question. It tells you who to hire — an engineer, a generalist, a contractor first. Or it hands you a feeling: you’ll know it’s time when you’re drowning. Neither is decision-grade. “You’ll know” is not a number, and the person who feels most like drowning is often the worst person to be making a hiring call. The question that actually has data behind it is narrower: what predicts whether the hire works out? And the honest answer, from the two best datasets available, is not timing. ...
There’s a question every solo builder answers eventually, usually by getting it wrong once: Which answers am I allowed to act on without checking? Not “which tasks can AI do” — that’s a capability question and the answers change every few months. This is a sorting question, and it doesn’t change with the model. It changes with the task. The framework — and why it isn’t the article Let me be honest up front, because we do this on every post: the grading framework itself is not new, and it is not ours. Regulators, enterprise security teams, and academic groups have all published versions of it. Microsoft’s agentic-risk guidance says require approval for high-risk or irreversible actions. A recent DeepMind paper on delegation names reversibility as the axis that decides how much authority you can hand over. Practitioners publish permission matrices grading actions by blast radius × reversibility × stakes. If you want the general framework, those sources will serve you better than a solo-builder blog, and I’ll list them at the bottom so you can check them. ...
A brokerage just gave every customer a button that trades for them. Over 150,000 people turned it on in four months. The agents hit the platform’s tools about 30 million times a day. That’s not a bug report. It’s a structural change, and it matters more than the headlines about “AI trading” suggest. Here’s the thing nobody says plainly: those 150,000 accounts aren’t running 150,000 independent strategies. They’re choosing from a handful of shared models — the same few AI systems, pointed at the same market data, optimizing toward roughly the same goals. ...
For most of the history of shipping software, a solo builder could answer one question with a single, static number: “what does this cost me a month to run?” You tallied a server, a database, a couple of SaaS subscriptions, rounded up, and you were done. The number was flat. You built to it and forgot it. That number is now fiction. The tools you build on stopped charging you a flat rate and started charging you by the meter. GitHub Copilot — the most common AI coding subscription on the planet — moved to usage-based billing on June 1, 2026: every plan now draws down monthly “AI Credits” pegged at one cent each, with the old automatic fallback to a cheaper model removed entirely. Anthropic now estimates the average developer burns $13 per active day in Claude Code token spend — up from $6, doubled in one quiet docs edit. Flat-rate is over, and the solo builder is the one who feels it hardest because there’s no CFO to absorb the surprise. You are the CFO, the engineer, and the one who pays. ...
The AI model is out. Not gone — out. Finished as a place to compete. Everyone still talks about the model as if it were the whole story: which one is smartest this week, whose benchmark just moved, who shipped the biggest one. None of that is the important thing anymore. Here’s what actually happened. Token prices fell roughly 600× in six years. A model that cost $60 per million tokens in late 2021 now costs pennies. Open-weight models do for free most of what the flagship ones charged for. The model became a commodity — something you rent by the meter and swap out whenever a cheaper one is smart enough. ...
AI makes building cheap. It does not make building valuable. Those are two different problems, and most builders only notice the first one. Give a modern AI tool a clear description and it will produce the feature — the research, the code, the tests, the build. Hand it ten rejections from an app store review and it will diagnose and fix each one. That part of the old job genuinely got faster. ...
“Solo founder” used to be a phrase investors said with a wince. It read as couldn’t find a co-founder — a red flag on a term sheet. Going it alone was the backup plan, not the point. That framing is now dead, and it didn’t die by argument. It died the way framings actually die: the data stopped supporting it. This post is about the two datasets that killed it — and about the one number in them that most people quoting these stats get backwards. ...