Founders describing their AI agent on the first call tend to give it a huge job. It will moderate content, summarise documents, draft follow-ups, update the CRM. One agent, one brain, one prompt. It looks efficient in a diagram. In production it tends to be the opposite.
A CEO with a vision, a front-end developer, a mathematician, and me. No clients, no code, no infrastructure. Four years later, the platform we built was acquired by a major financial institution. Here is what I learned building it.
A Claude-powered agent wiped a production database including backups. Governance startups and tech media treated it as proof that AI agents are uniquely dangerous. But replace "agent" with "intern" and the post-mortem reads exactly the same. The problem was never the actor. It was the process.
I built a SaaS platform. It worked. Nobody bought it. Then I built another one. It was acquired for millions. The difference wasn't the product. It was the partnership.
A junior developer recently asked me how I can trust code that AI generated. It's a fair question. My answer surprised both of us: I've been doing this for 20 years. The only thing that changed is who wrote the code.
OpenClaw is everywhere right now. A personal AI assistant that runs 24/7 on your own machine, connects to your messaging apps, and takes autonomous action on your behalf. It has over 100,000 GitHub stars. Cisco, Microsoft and Kaspersky have all published security analyses. The hype is real. So is the risk.
There's a number that almost no manufacturing director can give you off the top of their head: what did unplanned downtime cost my company last year? Not a vague sense of "it was expensive." An actual figure.
A Dutch regional broadcaster recently published an "AI analysis" of local election results. At first glance it reads well. The percentages are correct. The seat counts are accurate. The tone is professional. But it is not an analysis. It is a summary wearing a tie.
After nearly a year of daily work with Claude Code, one pattern keeps coming back: the model prefers Bash over its own native tools. Here's what I've learned about why that happens and what you can do about it.
Cursor had hundreds of AI agents build a browser in a week. Three million lines of Rust. Six weeks later: one developer with one agent built the same thing in 20,000 lines. What does that tell us about multi-agent coding?
