Most teams don't need more AI agents. They need an org chart for the ones they already have.

I keep seeing teams talk about agents like the next step is just adding more of them. I think the next problem is simpler and uglier: agent sprawl. Not just too many agents. Too many overlapping permissions and quiet little workflows nobody remembers until some other team gets hit by something weird. The hard question is not "can we get an agent to do this?" It's "what happens when six of them do related things across the same stack all day and nobody owns the full picture?" If I had to put a basic control layer around agents inside a company, every serious agent would need five boring fields: owner, systems it can read, systems it can write, a budget or usage cap, and a stop rule for when it has to hand off to a human. I'd also split them into four classes: readers, routers, operators, and spenders. A reader is not a spender. A router is different from an operator. If you treat them all like one blob called "AI agents," you either over-control harmless stuff or under-control the expensive stuff. Curious how other people are handling this. Do you actually keep an inventory of your agents anywhere yet, or is most of this still living in people's heads and scattered docs?