What the heck is an AI Agent, anyway?

🧵 An April 2025 working journal from someone stumbling upward through the “year of agents.”

1. Let’s Start with the Classic Definition

“An AI agent does something on your behalf.”

Great. Thanks. Super helpful.

I always thought the laziest form of writing or public speaking was "Webster's Dictionary defines this as…." And that's about where we are on this explanation of AI agents. It technically includes everything from a basic chatbot to a full-blown autonomous system trying to rewrite your OKRs.

So yeah, something on your behalf. But what? And how? And can I trust it?

2. The Digital Teammate Model

Here’s where things get spicy.

An agent isn’t just a tool—it’s a co-worker. A digital teammate.

They join your Slack. They remember past work. They write updates. They don’t (ideally) hallucinate your Q2 forecast. They learn. They report.

Will an AI agent literally sit in your org chart? Experts say it's a matter of when, not if.

This is where most SMB leaders start getting intrigued and a little nervous. Because if your AI teammate has memory and autonomy, that means you’re managing them now.

3. Entities With Deep Memory

A real agent doesn’t reset between tasks like a goldfish.

It remembers:

• What you asked it yesterday

• What worked (and what failed)

• Who is on the team and what are their tendencies

• The overall goal of the week, month, quarter, and year

In other words: it builds context and doesn't reset after each session. That memory makes it exponentially more valuable the more you invest in it. If you had an employee that never forgot anything, how valuable would they be? I'm genuinely curious.

4. Work-as-a-Service vs. Results-as-a-Service

Here’s where the org chart gets blurry.

If you’re hiring an agent…
Who manages them?
How many can one person oversee?
Do you need a Director of Agents by Q4?

The SaaS world has taught us to think in units of software. The agent world flips that: now you hire for outcomes, not just features. Enter: Work-as-a-Service (WaaS) and Results-as-a-Service (RaaS).

I’ve started sketching these as two future business models:

• WaaS = “The agent will handle it.”

• RaaS = “The result will appear.”

Still wrapping my head around this one—but it’s already reshaping how I think about staffing, automation, and accountability.

5. Atomic Agents (h/t @dharmesh)

This is one of my favorite models so far:
The Lego Bricks of AI.

Each agent does one job. Very well.
They’re stackable. Reliable. Less likely to go off the rails.

Think: one agent to summarize. One to rewrite. One to route to the right team.

You can compose workflows from these atomic pieces—just like you’d build a team of specialists instead of hiring one chaotic super-generalist.

Also: the James Clear fan in me is very into the atomic metaphor.

6. The IQ Threshold

When AI can perform at or above the 90th percentile of every knowledge domain, then we'll be in the Agent Era. That's a common belief among thinkers in the space.

The models are already there in some fields (Harvard Business Review would be happy to tell you about it).

7. AI as a Scientist

Another definition I've heard is that "we'll know we've reached Agentic AI when the AI can make novel scientific discoveries." I searched far and wide for the single person to attribute this to, and didn't have a ton of luck. Want to read a paper on it?
“What if you team up a bunch of agents?”

I really wanted to make a joke about AIsaac Newton here, but even I cringed typing that out. Leading researchers in the health space say we aren't far from this, by the way.

So what is it?!

Everyone says “agents are the future.”
Cool. Ask them what they mean by “agent.”

Better yet—ask what their org is actually doing with them.

As for me? I’m still deep in the exploration zone.
Building, breaking, learning, logging.

This is my journal, and you’re welcome to steal what works.

🛠 Got a working definition of "agent"? I’d love to see it. Drop me a note or tag @ceoAIcoach.