Have you ever asked a question to a room full of first graders?
That is basically how my first meeting with AI agents went.
I asked a question and expected a team discussion.
Instead, both agents started talking at the same time.
Same topic. Same moment. Both confident. Neither pausing. Neither asking the other what they thought. Just two voices crowding the same lane, like nobody had told them they were sitting at the same table.
I laughed, because I'd seen this meeting before.
In 2020, I started building a global team. COVID had shut everything down, and I needed people I'd never met, in countries I couldn't fly to, doing work I couldn't watch happen.
One of my first hires was my Amazon guy. I asked him to audit our listings: check search terms, make sure we were using what we needed to rank organically, clean up the basics.
He was sharp. Reliable. Cared about the work.
And then he froze.
Not because the job was too hard. Because nobody had told him what done looked like.
So he kept coming back.
Is this listing finished? Should I include this variant? Do we count long-tail keywords? What about back-end search terms? How many is enough?
Every question was reasonable. Every question was a sign that the lane was undefined.
I could have called him slow. I could have called him uncertain. Both would have been wrong. The problem wasn't him. The problem was that I had handed him a task without handing him the definition of done.
That's not a people problem. That's a job design problem.
I learned that lesson again and again building the team. The number one reason my people failed wasn't talent. It was loose job descriptions and undefined finish lines.
The fix was never a motivational speech.
It was structure.
So when I watched my two agents talk over each other, I wasn't frustrated. I knew exactly what the problem was and how to solve it.
This wasn't a prompt problem. It wasn't a model problem.
The lanes were unclear. The finish lines were undefined. The handoff rules didn't exist.
And that was on me.
Here's the takeaway for my fellow operators.
Don't deploy your AI agents like you're implementing software.
Build them like you're creating the perfect employee.
Agents don't sit. They act. They send emails, update records, query vendors, summarize invoices, and hand work to other systems.
Without defined roles and boundaries, you don't have an AI workforce.
You have automated chaos.
You don't give every employee the same access. You don't let a new hire email a major account without knowing the rules. You don't ask a sales coordinator to approve freight claims.
That's not distrust.
It's scope control.
Treat your agents like digital employees.
Same playbook. New kind of worker.
First, define the role. Chief of Staff. Executive Assistant. Research Analyst.
Second, define done. If the finish line is unclear, the agent will improvise.
Third, define the process. What happens first? What gets checked? What happens when something doesn't fit?
Fourth, set the guardrails. What can it access? What requires approval? When in doubt: read-only.
Fifth, define escalation. Stop and ask? Hand off? Wait for you? Silence is not an escalation path.
Five things.
Role. Definition of done. Process. Guardrails. Escalation.
The same five principles, just applied to a different kind of worker.
You're not deploying software.
You're assigning work.
The next time one of your agents drifts or freezes, don't blame the model. Don't write a longer prompt. Don't bolt on another tool.
Pull out a blank page and write the job description you forgot to give them.
That's the work. The solution is right in front of you.
What's the first agent role you'd define in your business?
Lock in and set your mind right.
Ricky