Ok, ok, ok, we're talking AI Agents

In last week's email I hinted at AI agents and then actively ruled it out as a topic.

I then received multiple replies asking why. Why is it not ready? Why won't you write about it? And the big one…

What is it?

This newsletter has been breaking the "3-minute weekly email" rule for a while now so it may be time to update the slogan.

But like the AI world, things change quickly, and we're an adaptable bunch so let's settle in for a long one.

What is Agentic AI and why is it not ready?

If you ask me what determines an agent, I use one word: autonomy.

This isn't your standard ChatGPT window waiting for instruction - agents are systems that can survey a situation, choose their approach, and act on their decisions.

At their core, agents can:

• Make independent decisions beyond their initial programming

• Select and operate tools without explicit instructions

• Adapt their approach based on previous success and failure

• Take initiative in problem-solving

The easiest way to visualize this is through an architecture diagram. I've created one based on a hypothetical personal healthcare agent that can:

• Collect data from your smartwatch and health records

• Perform medical research on the latest advances in medicine

• Analyze your information against what it finds

• Create a plan and notify you of any early warning signs

(Larger image here)

I don't know about you, but diagrams like that excite me. They show us the potential of truly autonomous systems that could revolutionize how we approach complex problems like health. Now I'm no medical expert, but you'd say a system like this guided by medical professionals would generate a decent amount of revenue too 🤔.

But here's where excitement meets reality.

Last week, when I mentioned AI agents, I was direct about my position: they're not ready for prime time and aren't meaningfully impacting our work lives... yet.

The technology - while promising - is simply inconsistent.

Current setups can demonstrate outstanding results in controlled environments but often struggle with real-world complexity. And applying it to healthcare? Well, I think we're a long way off.

But this gap between agentic potential and human application isn't an insurmountable hurdle - it's a signal that we're in a crucial developmental phase, kind of like when ChatGPT was released all those years ago.

What Agentic AI is not (and why the alternative is still fantastic)

If you look up "building AI agents" on YouTube, you'll find countless videos showing how to set up agents in Make.com and Zapier Agents.

To be blunt: these systems are not AI agents - they're AI workflows.

AI workflows are deterministic - the path is "determined" by the creator of the workflow. Given the same input, the system will always take the same steps and produce the programmed desired output.

Agentic systems, by contrast, are inherently indeterministic. They might take different approaches to the same problem based on their current understanding or context. This unpredictability is the feature that enables genuine problem-solving and adaptation.

You'll see this indeterministic nature perfectly illustrated in Sam Witteveen's video - when asked to research and write an article, the agent didn't follow a predetermined path. Instead, it independently assessed the task and chose its own approach, so much so that it didn't even consider stopping to ask for human input. You couldn't predict exactly how it would tackle the task - you could only observe as it made its own choices and moved forward. Sam then had to build a customer interaction agent specifically to ensure the system would remember to include human input in the process.

And this is precisely why businesses prefer deterministic workflows: they're reliable, predictable, and easier to debug when something goes wrong.

And so why doesn't this distinction matter for us?

Simply: workflows can be incredibly powerful even if they aren't true agents.

With AI workflows, you can:

• Process incoming emails and draft replies based on predefined criteria

• Analyze customer sentiment across helpdesk tickets

• Monitor news feeds and match companies' press releases with active job postings on Indeed

The key isn't whether a system is "truly" agentic - it's whether it solves your problem effectively.

Homework

I juggled with this week's homework - do I demonstrate how to get stuck into the complex world of agents or do I go with workflows? I decided that demonstrating workflows would be a disservice to those who asked for a chat about agents. I found the easiest set up I could find that was powerful enough to give you an idea of where this is all headed- so let's get into it.

We're going to set up a Meeting Preparation Crew using Crew AI's pre-built templates. This isn't just another workflow - it's a team of researchers who will prep you for a meeting. In this case it’s a meeting with Mark Zuckerberg (we’re going to need more agents!).

1. Go to https://app.crewai.com/ and sign-up (it's free)

2. Go to https://platform.openai.com/api-keys and create an API Key (this is not free but very cheap - just monitor your usage on this page)

   Go to https://serper.dev/api-key (this is free for a limited amount of tokens)

3. Click on templates in the side bar, find the "Meeting Preparation Crew" and click "Deploy"

   

4. Enter your two API keys from steps 2 and 3

   

5. Hit “Deploy” - this process will take around 5-10 minutes (you’ll receive an email once it’s complete)

6. Once deployed, select “Crews” from the side menu and select “Manage” on your new crew

   

   As of

7. Once your crew loads, select “Trigger Crew”, fill in the details of what you want to achieve and then select “Trigger Crew” again

   

8. Once completed, you will see a job like this in the “Completed” column

   

9. Click on the “Output” dropdown to see your results

The bottom line

Sheesh, that was a lot. It was way more technical than we usually go yet even with the complexity, we only scratched the surface of AI agents.

But this distinction between workflows and true agents matters. It matters because right now, we're standing at an interesting crossroads. On one side, we have incredibly useful workflow tools that can automate real business processes today. On the other, we have the emerging world of autonomous agents that hint at a future where AI systems can truly think and act independently.

Understanding this difference doesn't just help you avoid the marketing hype - it helps you make better decisions about what to implement now and what to watch for in the future.

Because while true agents might not be ready for prime time, they're coming.

And when they do arrive, you'll be ready to stare them straight in the eyes - if they have eyes.

IN PARTNERSHIP WITH THE RUNDOWN AI

Learn how to make AI work for you

AI won’t take your job, but a person using AI might. That’s why 1,000,000+ professionals read The Rundown AI – the free newsletter that keeps you updated on the latest AI news and teaches you how to use it in just 5 minutes a day.

Sign up to start learning.

FACTS TO IMPRESS PEOPLE AT PARTIES

Unlike regular AI systems that need human guidance for every task, autonomous AI agents use a principle called 'recursive self-improvement' - they can actually analyze their own performance, identify weaknesses, and modify their approach without human intervention.

It's essentially AI teaching itself to be better at its job… Is this getting scary yet?