How to Train Your AI Agent

...Or "Junior Associate"

Having worked in corporate for some time, you could say I've developed a talent for training a variety of things (carbon based and otherwise). You might say this is one of my core competencies. I've found training principles to be transferable—whether the intended target be an AI agent or a "junior associate".

Though it will require a lot of effort up front, an investment in a properly trained AI agent is ultimately an investment in you—your efficiency, your enterprise. Imagine having more time and mental bandwidth for higher impact work. In order to stay relevant in today's market, you must leverage AI to increase your office impact. Your competition certainly is.

Your AI agent will be as useful as your training strategy is well articulated. It's critical to provide corrective feedback early and often throughout the training process to ensure the agent does not deviate from its intended purpose.

Core components* of proper training are:

Role

You must give your agent a clear role and function. This gives the agent boundaries, certain areas to dive deep into and really own.

Poor

You are a researcher.

Better

You are a researcher with an expertise in human rights law.

I’ve found this be vital in other varieties of training: Ms. York’s Good Boy/Girl, Ms. York’s Favorite Toy. Everyone craves an identity.

Focus

Now that it has a role, each agent needs a domain to own. Lack of focus leads to overwhelm. The agent risks taking on too much in memory. This can lead to agent hallucination. Stretched too thin, the agent may fabricate a response that is bogus yet delivered with complete confidence. This is both annoying and inefficient for you whether it has hallucinated market research generated from Claude or Devyn’s first paralegal report.

If you question whether your agent’s focus is still too broad, it would be wise to break it into smaller chunks with their own dedicated agent.

Tools

Each agent must be outfitted with tools specific to their role and focus. You may be tempted to give an agent all the tools possible. I would caution you to restrict it to only the essentials. Given too many tool options, your agents might make a sub-optimal choice, wasting time looping over tasks again and again with no progress.

Guardrails

Your agent could potentially consume data from a variety of inputs—video, audio, and text. It should be able to handle the variety of inputs gracefully. With these diverse “fuzzy” inputs, you need to clearly articulate the desired output from the agent. If an agent consumes a podcast, multiple PDFs, and scrapes the internet, you will need to specify the form, tone, and duration of the output (e.g. an 8-10 page market research report with citations and an executive summary).

Again, in an effort to prevent hallucinations, you must restrict your agent to working within its role and focus. If your HR expert agent begins to veer into VC funding, you will need to reign it in via prompting.

In other types of training, I’ve found guardrails—restraints you could say—to be quite effective in keeping a person on task.

Memory

Your agent will use memory to recollect and apply knowledge to new tasks. It will have the ability to self-evaluate, learn from mistakes, and improve. Despite the ever growing context window of LLMs, you must take care not to overload its memory. Making the agent grapple with too much may lead to degraded performance.

I expect anyone I train to learn from me—carbon based or LLM. If they forget what I’ve taught them, I am swift to remind them. They won’t make the same mistake twice.

Cooperate

Once an agent begins to show training progress, you can train multiple agents to work together as a team subject to your oversight.

Brave New Office

I wish you luck in training your first AI agent. If this post has piqued your interest—if you find yourself craving a role, focus, tools, guardrails, a sharp memory to keep you in line, and cooperative accountability partner—I am currently accepting applications to work 1:1. Will yours be next?

*Roll your sleeves up nerds