17 posts
Episode 5: six shutters deciding on their own, from their room temperature and the sun's track across their facade. The model was built from the land registry and thirty days of correlation. Its most useful output was telling me where not to act.
Episode 4: fifteen spoken announcements on a school day, from wake-up to bedtime. The sound played fine from day one, and nobody moved. Eight writing rules, a wake-up rewritten three times in one day, and what happens when the bedroom speaker is offline.
Talking houses, a smart car, air taxis, humanoid robots, brain-computer implants: the films of our childhood are meeting the present. A tour of the works that came true, and of what this timeline still costs.
Episode 3: the house now carries the shape of the day. Gate open at 7:55, television cut at 7:30 p.m., shutters closed at the same moment, and an official open dataset that suspends the whole thing during school holidays. Every piece took an evening, and that is what made changing my mind three times possible.
Over a year of home automation run through an agent driving the Home Assistant API: 87 devices, nine ecosystems, 75 automations. What I got out of it is ideas I would never have tried, and diagnostics that cross five systems in one sitting.
A survey of a ten-year home automation build: from Jeedom and hand-flashed ESP8266 boards to 87 devices across nine ecosystems. What I built alone, what arrived once I started working with an agent, and the two examples that carry the difference.
Product ownership, testing, security, platform engineering: most of what got hired as a separate seat over the last ten years is a skill that never needed one. On my team the fix landed on the tech lead, who now carries product judgment next to the technical expertise he already had. One field report, not a template — and it updates something I only got half right two posts ago.
Yes, I trust the frontier models — I run them every day and I know what they produce. But calibrated confidence is still personal, not proof about the system. Here is the quality system I built on a regulated hospital ERP: what quality means when the software has to hold up in court, the 4 layers that enforce it, the ladder we measure it on, and what it still does not prove.
The same week, my feed buried the Product Owner, saved the Product Manager, gave up reading AI code line by line, and handed a whole product to one person. Four posts, one event: scarcity changed sides. What dies is the relay function. What survives is product judgment — and it no longer belongs to a single role.
Eight frontier models in six months, treated as what they have become: a versioned dependency that depreciates. What compounds is the doctrine — the ADRs agents read, the rules CI enforces, the versioned context. And a confession: our own inventory found the doctrine drifting — the part nothing executes had started to lie.
Agentic AI removed the friction between an idea and running code — and the friction was the only brake I had. I build seven days a week; roughly 90% of what crosses my mind becomes a project. A first-person account of the hook: what behavioral science actually says, the sourced ADHD/hyperfocus detour, and the two questions that hurt — does any of it ship, and is it worth what it costs?
Eight frontier models in six months, most sold on self-verification. On Solstice, our greenfield platform, four months of development with redefined working methods turned model quality into velocity: roughly 4,500 commits, close to 1,000 PRs, 120-plus releases. What we automate in July — and what it took to get there.
Google Cloud's DORA team released its 2026 report on the ROI of AI-assisted software development. The J-Curve, the verification tax, AI as an amplifier — a field report on research that validates what my team lives daily, vendor numbers and all.
Six months into 2026, my team's workflow has quietly rewritten itself — and the job has never been this enjoyable. A panel of persona sub-agents reviews every PR and every issue. Meetings record themselves and turn into assigned GitHub/GitLab issues. The boring parts are gone. What's left is business, domain, product. The work we always wanted.
OpenAI just launched a $100/month plan between Plus and Pro. We are silently building a two-speed AI market: the cheap AI everyone comments on, and the premium AI nobody knows. The public debate is biased at the source.
Coding was never the job. It was the bottleneck we mistook for the craft. AI removed it. What remains is what always mattered: architecture, domain expertise, product vision, and judgment. A look at the new roles emerging in software development.
A point of view on the state of software development in 2026. Why the SDLC was always broken, why AI agents are making it impossible to ignore, and why European tech needs to wake up.