Metorial blog posts about opinion.
There is no meeting where someone decides to build an internal AI platform. It just accretes, one reasonable decision at a time, until eighteen months later you have thirty-five thousand lines of glue code, one person who understands it, and no documentation. Here is how that happens, and why the fix is not writing better glue.
I built an internal agent to clean up our CRM. It worked great — until a friend asked me what it was actually authorized to do. When I went and looked, I realized I had handed an autonomous process admin credentials with no expiry, no scoping, and no audit log. This is that story, written down plainly.
We’re in the jQuery age of AI agents, where every integration feels like duct tape, and standards are still a dream. Just like jQuery once smoothed over the chaos of incompatible browsers, MCP is doing the same for fragmented AI tools, making it finally possible to ship agents that actually work. It’s not perfect, but it’s the layer we need right now.