De-Mystifying AI Agents – Sharing 3 Lessons Learned

That's me, writing about the concept of artificial general intelligence (AGI) in 2018.

Eight years ago seems like ancient history! By November of 2022, the general public could use ChatGPT — and in the 4 years since then, there's been a monster wave of activity. We're still not at the definition of AGI I learned about back then. However, what's possible with the latest tools like Claude or ChatGPT continues to accelerate fast.

As I did back in 2018, I decided to go deeper to understand one of the latest areas of AI that one hears a lot about – AI agents.

I decided to write about it again with the hope that I can add a voice of clarity and simplicity to others who are navigating this wave. I feel like in 2018 I was slightly ahead of the big wave, and now I'm chasing it — but I hope this can still help some people in my professional network feel less overwhelmed and better able to tackle the next opportunity in this space.

Lesson Learned #1 – Agents range from simple to complex

A large language model (LLM) like Claude predicts the next word, one at a time. Ask it to fix your grammar and it does a nice job — then it's finished.

An agent is different. It has a goal, it can reach outside itself for information, and it can connect to and use other software tools. It keeps going until the goal is met.

The instructor of my Stanford continuing studies course on AI agents described agents as living on a spectrum — a spectrum of degrees of freedom to act on its own. On one end of the spectrum is one of the latest LLM's answering questions you ask it in a chat, using data it has in its trained memory banks. In the middle, you can build an agent that automatically gives you a daily briefing every morning by scanning your email, calendar and Slack — and continuously improves itself. On the far end, you break a process down into multiple tasks and create agents for several of those tasks — with those agents independently acting, sharing information with each other, and coordinating to complete a more complex workflow. There's no clean line between "not an agent" and "agent." You add capability one notch at a time.

This framing helped me more than any definition ever did. The question stops being is this an agent? and becomes how much independence am I willing to give this agent to take actions on my behalf? Will you let it book a flight, pay for it with your credit card, and put it on your calendar? Will you let it decide on its own whether to give a customer a refund?

Lesson Learned #2 - You don't need to code. You still need to plan and think.

I have a client with an employee who came from a background as a medical assistant, doing benefits access reviews. No formal training in technology. She taught herself to build simple agents using tools like Claude and has since automated several tasks — for her own role, and for others across the company. Her contributions are reshaping her role, and how she's paid.

You can build a working agent quickly without writing a line of code. You attach the files it should read. You connect it to tools like Slack or your calendar. You write down the steps you want it to follow. You set limits on what it's allowed to do. It all seems simple to do.

But these basic AI agent creation steps are not the hard part. The planning and thinking are.

Good agent design — especially once a workflow gets complex — leans on disciplines that are a lot older than AI. Process design. Thinking deliberately about how a technology should serve a goal, instead of bolting it onto whateveryou're already doing and hoping for the best.

There's a great test from the course: could you onboard an intern to do this? Could you explain every step, define what done looks like, and hand them every tool they need? If you can't do that, an agent will fail at it too.

Four things you must supply to the AI agent:

• ‍Knowledge — What it needs to know. Policies, documents, past examples.‍
• Tools — What it can reach. A calendar, a database, a travel booking system, a payment system.
• Skills — How it does the recurring work. The steps, the standard, the format.
• Guardrails — What it must never do. Draft the email, don't send it. Never post without approval.

Logging and evaluating the actions of an AI agent matter just as much as the rest — more, once you give the agent more independence. In 2024, a Canadian tribunal ruled that Air Canada had to honor a bereavement-fare discount its customer-service chatbot invented out of thin air. Building in a loop of detailed logging, evaluation, feedback, and improvement is how you catch that kind of thing before a customer does it for you.

Lesson Learned #3 – This is also process and org design work

In multiple forums lately, I'm seeing a checklist for using AI agents that looks something like this:

• What outcome do you want?
• What's the current workflow — and is it even the right one? Fix that first, or you'll just automate a bad process faster.
• For each decision in the steps of the workflow: human or agent?
• What does it need to know to decide well?
• What does it do, and what does it hand off to whom?
• Who does it escalate to when conditions get strange?
• How will you know it's working?

This isn't new thinking. It's the same discipline behind any good process design or automation effort — it just happens to be running on a different kind of engine this time, one that is incredibly powerful and allows you to build and deploy much, much faster.

Implications for Senior Leaders and HR Professionals

My professional network on LinkedIn is heavily populated by senior leaders and HR professionals. For any of those folks who might be feeling like I did a few months ago, the good news is that it's not hard to cut through to the heart of it all.

The scarce skill here isn't prompting. It's knowing how work gets done, where the decisions live, and what must stay human. You do that for a living. Continue working across the organization to find opportunities to apply these new technologies for the benefit of your customers and employees.

In 2018 I wrote that AI was another wave of automation. That part held up. What changed is that the waves are bigger and faster, but the average person no longer needs software coding skills to ride it.

(Disclosure: I had Claude draft the first version of this using my notes from multiple courses and webinars. It was the humans who made the difference: the instructors at Stanford continuing studies, multiple speakers on multiple webinars, my friends who gave me feedback. Thanks to all those people!)

‍