On agent swarms

The limit we keep hitting is how many decisions per hour a system can actually make. Every agent run we build stalls at the same kind of step, the one that needs a judgment call, with a human somewhere who has to supply it. One agent with one person watching it is fine. A hundred agents running in parallel need a hundred of those judgment calls at the same time, and that's what our workload looks like now. One person can service a handful of those calls an hour, and a fleet asks for far more than that at once. Waiting on a person for each decision already looks impossible, so we have to scale decision throughput drastically. The way out is agents controlling other agents. A supervising agent watches its fleet, reads state, kills bad branches, reallocates work, and answers the questions that would otherwise queue in someone's inbox. Humans stay in the picture for direction, budget, and the expensive or irreversible calls, but we come off the critical path of every small choice, because…