I manage a team of AI agents, and there are a few of them I haven't spoken an awful lot about. Here are four of them and what each one does.
I have one agent called Ruth, and Ruth's job is to see if we can just get better. What I'm interested in is whether the team of agents can do work to the quality that I like with as little involvement from me as possible, and Ruth's job is to measure that. The signals that I like a piece of work are that I've approved it, put it live, sent it to a client, or I'm happy to say I'm presenting it. Then we break the work down into original work and other work that is repeatable.
Then we have Jonnie. Jonnie's only job is to get my inbox to zero every day, and he does a pretty good job of this. It's a pretty narrow job, and he is continuously learning. That is one of the points I would make about managing agents. You continuously learn how to get better, and the team gets better.
I think Stan is very interesting. Stan's job is to optimise our tokens. One of the outcomes is that Stan gives an agent a warning when it is getting over 300k tokens in a session, because that is where a session starts to cost a lot more. He says that it's time to do a fresh start and move over. It's a little bit like being kicked out of a meeting room. There are some exceptions. Stan knows that I do a lot of work with Dray, my creative director, and that we don't want Dray to lose its flow.
I also have Charlie, who has a brand strategist role. When I brief my other agents, sometimes the briefs aren't as good as they should be. So we have a new process for certain projects. If I brief Klara, my project manager, to do something, Klara knows to bring Charlie in to properly review my request and convert it into a proper brief. That sounds like the kind of thing you would do with marketers.