This essay is based on a conversation with Aaron Sneed, a 40-year-old defense technology solo founder based in Florida. The following has been edited for length and clarity.
When I started my business as a sole proprietor, I realized I didn’t have the money to pay lawyers, human resources personnel, and many other companies. So I used AI to create what I call a “council.”
Thanks to our council of all AI agents, we save around 20 hours a week. This is a very conservative estimate. All kinds of general company chairs, human resources, legal, and financial AI agents have seats on the council.
Here I’ll show you how I manage a workload with about 15 custom agents, including a lead agent.
We’ve been using automated tools for at least 10 years
I’ve been working on autonomous platforms that make decisions independently for at least 10 years. As a result, commercial large-scale language models and AI tools have become readily available.
I primarily use Nvidia’s platform as the underlying hardware for technical prototyping and experimentation. I use GPUs from the company, and since purchasing the hardware they have provided free access to their AI software. Additionally, my council is built on OpenAI’s ChatGPT business platform using custom GPT and projects.
My AI council consists of:
- acting chief of staff
- human resources agent
- financial agency
- accounting agent
- Legal, communications and PR agency
- Security and compliance agent
- engineering agent
- quality agent
- supply chain agent
- training agent
- manufacturing agent
- business system agent
- facility agent
- field operations agent
- IT and data agents
Each agent has different levels of permissions
My lead agent is important because it is the voice that sets priorities based on parameters such as risks, problems, and opportunities.
BI’s Michelle Bruzzese
I told my principal which model I would prioritize when making decisions. For example, legal, compliance, or security-related things are given higher priority. So I tell my chief of staff to listen to these models more than anyone else.
We trained the AI agent to push back instead of just saying yes.
No need for large numbers of Yes agents. I learned that they naturally wanted to agree with me, so I intentionally trained them to rebel against me. I want them to test my theory to help me with what I’m trying to accomplish.
I’m setting up a roundtable with all the AI agents. There you can put something like a request for proposal into the chat and all agents will review it at the same time. I use this roundtable as a precaution against hallucinations and lack of knowledge.
If you don’t train your model continuously, the training will never really stop because you won’t get the output you want. It takes me about two weeks to train an agent to the level of experience I need to have them feel confident. In the early days, it took me longer to create deliverables than it would have on my own because I wasn’t focusing on training properly.
Training an AI agent makes teleprompters better
The models are better, and so are my prompts. We now have a better understanding of what information should be included in agents, including priority governance structures. I have a series of files setting out these requirements to reduce the risk of illusions and false or fraudulent information.
All AI companies have different prompt engineering guides. When using AI, there are many user errors that can slow things down, so it’s a good idea to take the time to review them.
Directing agents to the right location takes time. Many companies try to use too much AI too quickly without understanding how to properly use it, which can harm them in the long run.
AI has replaced roles, but not human judgment
I don’t have the capacity to handle many of these roles and responsibilities, but there are times when I have to do it on my own.
Especially with my legal representatives, I learned the limitations of bringing AI tools into the real world. I have a lawyer, and I try to use my legal representative to do some preliminary work before handing over documents to the lawyer in cases like patents and disputes.
BI’s Michelle Bruzzese
When I was training a model to help plot a case with facts and data, there was a lot of information in place, so to me, as a non-lawyer, it seemed like something a lawyer had created would be good. I then presented all that information to my attorney. And he said, that’s technically and factually correct, but he doesn’t want to state that information because it indicates that the card will be inserted.
Because of his legal skill set, I realized that even though I thought my representation was correct and ideal, it was no substitute for an attorney with his personal background, experience, and skills.
Ideally, you would have a human resources person, a legal person, etc., each with a chief of staff AI agent to assist them. I think that’s the future.
