Understanding AI Agents (Without The Hype)

Today’s edition of the Augmented Advantage is co-authored by Dr. Tristan Behrens, hands-on AI advisor from Germany a.k.a. the AI Guru.

Hi there,

Another year is coming to an end, and with it comes a new buzzword in AI: "Intelligent Agents." If you're active on tech socials or follow AI news, you've probably noticed this term popping up everywhere lately.

But unlike many AI buzzwords that come and go, this one actually marks an important shift in how we think about AI.

In today’s newsletter, I’ve teamed up with AI Guru Tristan Behrens (who actually wrote his PhD on multi-agent systems over 10 years ago) to give you a realistic perspective on this topic and explore how your business might (or might not be) affected.

Let’s dive in!

What Are AI Agents, Really?

The term "Artificial Intelligence" has always been problematic, starting with its very definition. What exactly is "intelligence"? After decades of debate, we're still not quite sure.

AI Agents offer a more practical approach. First seriously explored by Stuart Russell and Peter Norvig in their seminal book "Artificial Intelligence - A Modern Approach," an agent is simply defined as an entity that can:

• Exist in an environment

• Perceive this environment

• Take actions to manipulate it

This environment could be our physical world, but it doesn't have to be. It might be digital spaces like the internet, databases, or even just your local computer. Think of it this way: both a humanoid robot navigating a warehouse and a software service monitoring your cloud infrastructure are agents – they're just operating in different environments.

The key difference from traditional AI approaches is that agents are judged solely on their performance metrics – concrete, measurable outcomes that tell us how well they're doing their job. 

For autonomous driving, this might be "miles driven without accidents". 

For a customer service agent, it could be "successful query resolutions".

Why Agents Matter Now

What's fascinating about 2024 is how people have started exploring the benefits of combining agents with Large Language Models.

Traditionally, agents relied on complex rule systems as their "brains." These complex rules have often been written in code and remained largely inaccessible to those without a computer science degree or a good degree of programming experience.

Today's agents are different. They use LLMs to orchestrate their thought processes, using natural language to:

• Gather knowledge

• Make decisions

• Plan next steps

• Learn from outcomes

Because these agents "think" in natural language, their reasoning becomes transparent to regular humans. No computer science degree required to understand why an agent made a particular choice!

Types of AI Agents

Just like with any technology, AI agents come in different shapes and sizes. They can range from simple reactive systems to complex learning entities:

1. Simple Reactive Agents

These are the most basic type – they just respond to their environment based on predefined rules. Think of a thermostat that turns the heat on when it's cold. Not very exciting, but they get the job done.

2. Planning Agents

These agents can think ahead and create sequences of actions. Imagine a delivery robot that not only knows how to navigate from A to B but can also plan alternative routes if it encounters obstacles.

3. Learning Agents

The most sophisticated type. These agents improve over time by learning from their experiences. This is where LLMs come in – they can help agents understand past interactions and make better decisions in the future.

Real-World Applications

Here's where it gets interesting for businesses. AI agents are already being used in various ways:

• Digital Assistants: Agents that can manage your calendar, email, and tasks, making decisions about priority and importance.

• Process Automation: Agents that handle complex workflows, like managing cloud infrastructure or coordinating software deployments.

• Customer Service: Agents that can understand customer queries, access relevant information, and provide coherent responses.

• Sales: Agents that can access CRM data to maintain customer records or help sales people work more effectively.

The Reality Check: What AI Agents Can (and Can't) Do

The key difference from traditional automation is that these AI agents don't just follow predefined scripts – which is both the main curse and the main blessing!

Like everything with "AI" in its name, expectations for AI agents are somewhat inflated. People expect them to magically solve tasks that neither traditional agents nor standalone AI solutions (like a simple LLM call) could handle. But let's be clear: agents are not a plug-and-play solution!

Instead, they need to be carefully designed and set up to work well within a specific scope. This often requires experimentation to find the sweet spot – a scope that's neither too broad nor too narrow to provide real value. And this is exactly where many projects stumble.

Remember the integration-automation framework? AI agents are just another item on your menu of potential solutions, alongside:

• Assistants (augmenting human work)

• Copilots (suggesting next steps)

• Autopilots (taking over specific tasks)

• Agents (orchestrating multiple steps)

The key is to understand that AI agents won't help you shortcut the iterative process of AI development. They're not a silver bullet – they're a tool that, when properly scoped and implemented, can help orchestrate complex workflows in a more transparent and adaptable way.

Conclusion

AI Agents represent an important evolution in how we build and deploy AI solutions. While they're not the magic solution some might hope for, they do offer a more transparent and adaptable way to orchestrate complex tasks – especially when combined with modern LLMs.

The key is to approach them realistically. Start small, experiment with well-defined scopes, and be prepared for the iterative process that any AI implementation requires. Most importantly, remember that agents are just one tool in your AI toolkit – choose them when they truly fit your use case, not just because they're the latest buzzword.

Speaking of that – in one of the next editions, this newsletter will probably dive deeper into some concrete examples of AI agents in action, showing you what worked and what didn't. Stay tuned!

See you next Friday, 

Tristan & Tobias