Latest
AI Best Practices Are Becoming Obsolete Faster Than They're Written
The playbook I was running six months ago is wrong now. Not outdated. Wrong. AI advice decays in months, and the decay is accelerating.
Read this postThinking on AI agents, accounting, and why the two are more connected than you'd expect.
AI is in production. That's settled. What most people haven't noticed: it stopped being a tool humans operate and started becoming a system that operates on its own.
Read this post
AI made production free. The one-person billion-dollar company still isn't coming — because we're probably only good for about 20 hours of judgment a week.
Read this post
For twenty years we treated skill as a definition. It was always a proxy for judgment. Then AI made the skill cheap — and exposed what we were really measuring all along.
Read this post
Job postings requiring AI skills in accounting jumped 67% in one year. What employers are actually listing reveals a profound confusion about what AI does — and that confusion has a price.
Read this post
POLA is 50 years old. Almost nobody applies it structurally to AI agents. Most frameworks hand the agent a god-mode API key and call it a day.
Read this post
Berkeley researchers studied 86 production AI agents. The teams that ship choose simple and controlled over clever and autonomous.
Read this post
Software companies are shipping AI agents that do your books. I went and looked at who's behind them.
Read this post
300,000 accountants left. Graduates hit a 20-year low. This isn't a hiring problem you can salary your way out of.
Read this post
Accounting solved the trusted insider problem centuries ago. One agent enters, one reviews, one posts. No single agent completes the cycle alone.
Read this post
A privacy policy is a promise. It can change with a board vote. Real security is architectural — the system is designed so misuse requires rebuilding it, not updating a document.
Read this post
Your agent's inbox is an attack surface. The fix: spawn a disposable subagent with no tools to read untrusted content. The intern opens the mail but can't wire money.
Read this post
Everyone building AI safety is thinking like an IT department. There's a better model — and it's been running since the Medicis.
Read this post
The AI reasoning layer changes every six months. The tools layer doesn't. Most AI companies are building the wrong one.
Read this post
Most agent builders stuff context windows with answers. The better approach is to give agents methods for finding answers.
Read this post