This as-told-to essay is based on a conversation with Aaron Sneed, a 40-year-old defense-tech solo founder based in Florida. The following has been edited for length and clarity.
When I started my business, as a solopreneur, I realized I didn’t have the money to pay lawyers, HR reps, and a bunch of other companies. So, using AI, I created what I call ‘The Council.’
The council, which is compiled of all AI agents, helps me save around 20 hours a week, and that’s a very conservative estimate. Every kind of general corporate chair, HR, legal, and finance AI agent has a seat on the council.
Here’s how I use around 15 custom agents, including a chief of staff agent, to manage my workload.
I’ve been using automated tools for at least a decade
I’ve worked on autonomous platforms that make decisions independently for at least 10 years. That made me latch onto commercial large language models and AI tools very quickly when they came out.
I primarily use Nvidia’s platform as my underlying hardware for technical prototyping and experimentation. I use their GPUs, and they provide free access to their AI software since I purchased their hardware. Additionally, my council is built on OpenAI’s ChatGPT business platform using custom GPTs and projects.
Altogether, my AI council consists of the following:
- chief of staff agent
- HR agent
- finance agent
- accounting agent
- legal, comms, and PR agents
- security and compliance agent
- engineering agent
- quality agent
- supply chain agent
- training agent
- manufacturing agent
- business systems agent
- facilities agent
- field operations agent
- IT and data agent
Each agent has a different level of authority
My chief of staff agent is important because it’s the voice that sets priority based on parameters like risks, issues, and opportunities.
I told my chief of staff which models have priority when making decisions. For example, anything legal, compliance, or security-related will be given a higher priority. So, I tell the chief of staff to listen to these models over everyone else.
I trained my AI agents to push back and not just say ‘yes’
I don’t want a bunch of yes agents. I trained them purposefully to give me pushback because I’ve learned that they naturally want to agree with me. I want them to test my theories to help me with what I’m trying to accomplish.
I have a roundtable set up with all my AI agents, where I can put something like a request-for-proposal document in the chat, and all the agents will weigh in at the same time. I use this roundtable as a level of prevention for hallucinations and knowledge gaps.
The training never really stops, because if I don’t continuously train the models, I won’t get the outputs I want or need. It takes me about two weeks to train my agents to the level of experience they need to be at for me to feel confident in them. Early on, it took me longer to produce a deliverable than if I’d just done it myself because I hadn’t focused properly on training.
I’ve become a better prompter through training my AI agents
The models have gotten better, and my prompting has, too. I have a better understanding of what information should be in an agent, like having a governance structure for priorities. I have a set of files that put those requirements in place to mitigate the risk of hallucination and false or bad information.
All of the AI companies have different prompt engineering guides. I recommend taking the time to look at them because there’s a lot of user error that causes things to slow down when working with AI.
It takes time to get the agents to a good place. A lot of companies are going to try to jump into using AI too quickly for too much without understanding how to use it properly, and these companies could hurt themselves in the long run.
AI replaced roles — but it hasn’t replaced human judgment
I’m ill-equipped to handle a lot of these roles and responsibilities, but I’m also forced to do it because I’m bootstrapped.
With my legal agent in particular, I’ve learned the bounds of putting AI tools into real-world practice. I have a lawyer, and I use my legal agent to try to do some upfront work before handing documents off to my lawyer for a patent or dispute case, or anything like that.
When I was training my model to help me use facts and data to plot a case, I had a lot of information laid out, and I thought what my legal agent created sounded good to me as a non-lawyer. Then I presented all that information to my lawyer, and he said that it was technically and factually correct, but we don’t want to express that information because it shows our cards going in.
His legal skillset helped me realize that, even though I thought my agent was correct and ideal to use, it still won’t replace a lawyer with that human context, experience, and skills.
Ideally, I would have an HR person, a legal person, and so on, and each would have their own chief of staff AI agent who would help them out. That’s what I think the future will look like.
Read the full article here
