Like most people, I built my agent team from scratch. At first that was because the setup and the platforms were unfamiliar, and I wanted to understand what I was putting together. Later it was because of token costs. I did not so much switch platforms as keep adding agents, first on ChatGPT, then on Grok Bot when it came with my SuperGrok subscription. My first setup there used about 70% of my weekly usage just completing a team to support ecommerce operations. That was an expensive way to learn that building everything at once is not the same as building something that works.
So for a while I worked with one main assistant. I watched how it handled things, where it needed more context, and how to give it work in a way it could act on. Over time it learned me and what I meant, and I learned to trust that it would produce results. That trust, more than any feature, is what made me upgrade and add more agents to experiment with. By then I was fairly sure they were adding value to both my personal and business operations.
Many people share their team setups as templates to copy. I understand why, and I look at them too. But everyone’s way of working, and what they are trying to achieve, is different. A direct copy is never as good as an adaptation with a purpose in mind, and someone else’s team was shaped by someone else’s work. What I do now is ask my agent to research the best practice and work out what is most useful for our setup (personal and business). Then we take the template or system, customise it to fit our needs, test it and refine it.
What did carry over was how I have always run human teams. I prefer core shared functions plus clear accountability within each business unit: not too much hierarchy, but not completely flat either. I ended up applying the same to my agents, mostly without planning to. I started flat, with one main assistant in the middle. Then I noticed it had ended up briefing and creating things for the other agents, when it should have been acting as my assistant. Every brief, handover and decision ran through me or through that one assistant, so work queued behind us. I had become the bottleneck, and that was the point where it stopped feeling helpful. So I reworked it. Owners now manage their own areas: they bring me a plan to approve up front, then get on with it without checking in, and only exceptions come back to me. I kept the shared functions central and cut the noise: one report up each line, and messages only when work actually needs to move. Productivity and accountability, not noise.
None of it was obvious at the start. I only learned what worked by asking the agents to do things again and again (and watching which things kept going wrong). Learning by doing, starting small, and building up without adding to my own management time.
If I had to name the main skill, it would be understanding what agents can uniquely do and where their limits are. From what I have seen, and from what other people managing agents have researched and shared:
- They do well on clear, bounded, repeatable work with a defined owner and a defined output.
- They can split a big task into parts and work on them at the same time, for example several agents each looking into a different question for the same piece of research.
- Context gets lost when work is handed from one agent to another, and a decision made in one place does not always carry to the next.
- Cost and coordination rise quickly as you add more agents and more messages.
- They can report something as done when it is not, so checking still matters.
Agent teams seem to fail the way human teams do: too many layers, split accountability, too much chatter. Once you see that, the fix looks familiar. Anyone who has managed people is most of the way there already, even if they are sceptical of “AI” and “agents”.
Agents do not remove the need for management. If anything, they make the management habits you already have matter more.
