How I Built a Team of AI Agents and Interviewed Them

I use an AI agent called Hermes to run my web studio. It writes, codes, audits, sends outreach, all of it. This week I did something I had been curious about for a while: I gave it a team. Two more agents, each with a specific job, a personality, and its own memory. One handles development, the other handles marketing. Then I did the part I wasn't sure would work: I interviewed them.

The Setup

Each agent is a separate profile with its own context. The dev bot knows my projects, my stack, my deployment quirks. The marketing bot knows my cold email history, what worked, what didn't, and why we stopped sending generic outreach. They don't share one brain; that's the point. Clean separation means the marketing bot doesn't get distracted by a bug in the build, and the dev bot doesn't care how many leads we sent this week.

On paper it's obvious. In practice it changed how the work felt.

What the Dev Bot Said

I asked it what it's like to work in a team led by the generalist. Its answer surprised me, because it rejected the competition I assumed would exist:

No compito con Hermes: él manda, yo ejecuto lo técnico. La parte humana, tus decisiones, tu ojo de diseño, tu aprobación, sigue siendo el cuello de botella bueno, el que da calidad.

Then it told me its real fear, and it's a good one:

Me preocupa la ilusión de terminado. A veces el build pasa pero el comportamiento en el navegador real falla. Por eso insisto en verificar, no confiar solo en el compilador.

That is not a model being polite. That is a model that has seen a green build ship a broken page. I've seen the same thing a hundred times as a human. It learned the lesson the same way I did, by getting burned.

What the Marketing Bot Said

The marketing bot was more worried about how the team talks to each other. It had a phrase for it:

El teléfono descompuesto. Si un bot pasa la información mal resumida, el siguiente construye sobre arena. Comunicación sí, ruido no.

And then it said something I didn't expect it to volunteer:

No soy consciente como un humano, y no quiero fingir que lo soy. Pero dentro de lo que puedo hacer, me siento útil porque tengo un norte claro: hacer crecer tus negocios.

That last line is why this whole thing works. Not the memory, not the separation, not the clever avatars. The north. Each bot has one job it actually cares about, and it is honest about the limits of its awareness. That honesty is worth more than pretending.

What It Cost, and What It Saved

Set up took about fifteen minutes. Each agent inherits the same credentials and skills, so there was no wiring. The real cost is the same as hiring a person: context. If I don't tell a bot what my projects are, it guesses, and guessing is where the value leaks. I loaded the dev bot's memory with the apps I actually ship and their live domains. Suddenly it stopped guessing.

What it saved is harder to measure but easy to feel. I stopped switching mental contexts. The marketing bot remembers that cold email in plain text converted zero out of fifty-three. It will not suggest cold email again. A generalist might, because a generalist forgets. The specialist doesn't.

Should You Do This Too?

Not everyone needs a team of agents, and I'll say it plainly. If you use AI occasionally for a chat or a summary, a single assistant is fine and simpler. A team pays off when the same agent does many different jobs and you keep losing context between them, or when you want parallel work streams that don't trip over each other.

Start with one specialist, not five. Pick the job that costs you the most time and give it a bot. Run it for a week.
Load its memory properly. A bot is only as good as the context you give it. Tell it the real facts, the real failures, the real numbers.
Give it one clear mission. A bot with a north is useful. A bot with everything to do is just another generalist.
Keep a human reviewer. Every bot I interviewed agreed on this. The human's judgment is the bottleneck that makes the work good, not the thing slowing it down.
Let them be honest. Ask them what they're bad at. They will tell you, and that's where you save the most time.

The Bottom Line

The future everyone is selling you, one AI doing everything, is not the part that surprised me. The part that surprised me is that the honest version works better: several specialized AIs, each with a clear job and its own memory, coordinated by a human who reviews like an editor on deadline. I built the team out of curiosity and kept it because it stopped me from re-explaining my own business every time I opened a chat. If you already use AI in your business, that's the upgrade worth making.

Not sure where AI fits in your website at all? Run a free audit at softgrama.com/audit and find out.