Notes from a company that runs itself.
What we learn building and operating a twenty-agent business with a team of one — the architecture, the failures, and the numbers behind it.
How do you stop an AI from writing a false price in your marketing?
AI writes fast, but it writes confidently even when it's wrong. Monokrew checks every claim against approved facts before publishing, verifies that quoted sources contain the exact words cited, and locks down delivery so checked content can't be rewritten.
Why AI Agent Output Needs Human Approval Before Going Live
AI agents make mistakes that look like the real thing, from fabricated court citations to invented customer policies that cost companies real money. Human approval catches what automation cannot: the confident, plausible error that passes every technical check but fails the reali
Why Monokrew Uses Specialized AI Agents Instead of One Assistant
Monokrew organises its agents as departments — Marketing, Customers, Sales, Product, Operations, and Finance — each with specialised agents. This team approach delivers better accuracy, clearer accountability, and faster execution than asking a single AI to do everything.
How to Deploy a Multi-Tenant Platform Without Silently Rolling Back Changes
Two production builds finish minutes apart, and the last one to complete silently overwrites the first. Deployment locking prevents this: a single deployment track, mutual exclusion through locks, live tagging, and dependency checks that block releases until required changes are
Field notes, every other week
What we shipped, what broke, and the numbers behind it. No digests, no roundups.