The Era of AI Agents is Here - Helping You Make Sense of it

AI agents are here. Well, it depends on how you define an AI agent. They are also still expensive, relatively slow, and often inaccurate. In most cases, they are also less autonomous than you might be led to think. However, they are already a powerful and critical part of the evolution of AI. 

Moreover, with the amount of investment by all the major AI players and many smaller companies and startups, we should assume exponential growth in the number and ability of AI agents in the next year. Thereby, all signs show that knowing how to build AI agents becomes the skill each company that strives to derive value from AI (a.k.a ALL COMPANIES) will need to master. The sooner, the better. 

In part 1 of a series of articles, I’m trying to go from hype to understanding.

Helping you learn everything you need to know about AI agents

I’ve spent the last two months immersing myself in the topic of AI agents. Primarily because we’re working on an agent for coding, I wanted to understand all the chatter and distinguish between hype and value. In my following articles, I will share practical suggestions and checklists for if and how to build an AI agent, tools to consider, and more. In this article, I want to share my understanding of an AI agent, the types of AI agents, their flow, and the different components and characteristics an agent has. 

Let’s start with a fundamental question: Is this a passing hype, or have we started the “agentic era?”

Agents are coming, in the masses, and this time for real. 

Agents are not a new concept. The idea of agents has been researched in the academia for many years. Until recently, they were not powered by generative AI, making them far less practical. Then, in the spring of 2023, with the introduction of AutoGPT and BabyAGI, we were all sent into a short-lived cycle of imagining an all-mighty know-it-all agent just around the corner. The hype over these early versions of generative AI-powered agents quickly subsided. They were just too broad and open-ended and, thereby, less applicable. 

This time around, things are different. We’ve come a long way, and with much more precise methods, bounds, and value-proving applications, it seems this is not another hype cycle but the true beginning of the agentic AI era. 

One might ask, but what is this “agentic AI era,” or even “what does everyone refer to when saying AI Agents?”

What are AI Agents? 

Experts agree that there is no single agreed-upon definition of an agent.

Each expert keeps saying: “Every expert you will talk to will give you a different definition of an agent..” and then they give their own quite similar definition. Most will agree with a broad definition: “Agents are software that can perform actions autonomously (to some extent - see the dedicated paragraph below), using some planning capabilities, with a clear goal in mind. Generative AI agents have at least one component that utilizes this technology.” The only caveat I want to mention is that in some cases, you might encounter abuses to this term for marketing reasons. That is when simple automation or straightforward use of an LLM is called an “agent.” Let’s put aside these non-agents for now and focus on agents that meet the above definition. When going beyond the generic definition of an AI agent, there is a significant variance in the content, complexity, and architecture of such an agent. 

Before we talk about the commonalities and differences between agents, let’s discuss their applications: 

What are agents good for? The main categories of AI agents and some notable use cases

The main appeal of AI agents is that they can perform more complex tasks than other AI or non-AI-based methods with less human supervision. As such, you can use them to automate work and outsource complex cognitive tasks. You can think about them as a “team of robotic coworkers” that can support your human ones or the personal assistant you never had. There are multiple types of agents one can think of. To help get a sense of the kind of use cases agents can be applied to, let’s start with a rough categorization of different types of agents:

1. Company representatives/ internal or external support agents - these agents will support your customers’ needs on your behalf. As such, they will be your digital identity, like company websites. The most prominent use case in this category is customer support. Notable companies working in this space include Salesforce, which rebranded itself as “Agentforce,” and Sierra, which focuses on customer-facing agents. Additional use cases include IT and HR support, which share similar requirements even though they cater to employees as customers vs. external customers. 

2. Persona-based internal agents - these agents emulate part of a specific role within a company. They can be comprehensive and complex, like “Devin, the AI software engineer,” or simple and mundane, like an agent used to automatically insert specific types of data into the company’s system of records. While coding agents are among this category's most promising use cases, there are also many additional use cases, like creating agents for finance and fraud analysis. Companies that support building agents in this category include vertical-specific agents, like cognition.ai, the creators of Devin, and horizontal agent-building platforms, including all the big players in AI that have already announced relevant offerings or are planning to announce soon. 

3. Personal assistants - These agents work on your behalf to save you time on repetitive, bureaucratic, or time-consuming tasks. You can find a combination of B2B and B2C use cases in this category. For example, arrange your calendar, book trips, or talk to customer support (agents?) on your behalf. 

Once we understand the main agent use case categories, we can start discussing how these agents work and how their characteristics differentiate what they can do and how. 

The main components of AI Agents and how they work

Don’t skip this part. Even if you plan on creating agents only using a friendly no-code drag-and-drop interface, you should understand an AI agent's various components and how it works. This will help you be more alert to the potential pitfalls and opportunities for improving the results. 

Below is my illustration of how a generic goal-based agent works: 

An illustration of a "classical" agent flow

1. A user or system request is actuating the agent. 

2. This actuation includes setting a goal by the user or defining one as part of the system that triggers the agent. The goal could be very open-ended and defined in natural language or from a closed set of known options. For example, a user-defined goal can be to “schedule a meeting with these people, book a room, and order refreshments.” A system-defined goal could be to “make sure this code is correct before adding it to the code repository.” 

3. The agent will then plan how to achieve the goal. This could be done using an LLM to help with the planning or through a pre-defined recipe for performing the task from a “Plan DB.”

4. Once the agent has a plan of tasks, it will execute them serially or in parallel. The execution step is where the agent does things “autonomously” to achieve its goals. For example, it might look up data, perform tasks on the computer, interact with humans, and more. Note that some tasks will be executed using AI while others will use other methods like automation or simple “if-else” algorithms. 

5. The planning and execution steps can use company or external data as a mandatory context or augment, ground, or customize them. 

6. The execution will be done iteratively with an analysis step where the agent assesses whether the single task or overall goals were achieved. If the goals are not met, improvements will be suggested. The analysis step will also apply guardrails and can decide to abort the agent. 

7. Once the overall goals are met, the results are sent to a human or integrated into the relevant systems. 

8. A powerful optional component of an agent is to have a “memory” of historical or relevant knowledge of how to operate. Agents that maintain a “memory” can make smarter, contextual decisions. 

 With this flow in mind, there will still be significant differences between various agents:

Some dimensions that characterize different agents: 

• Use case and domain - will impact the required knowledge, tools, properties, complexity, and more. 

• Business process complexity: Some agents will perform a business process with only a handful of steps, while others could perform processes comprising hundreds of steps. 

• Data sources: The data sources to which the agent has access when it was built or operates online will have a crucial impact on its behavior. For example, if an agent has a good representation of your company's inner workings, such as business processes, roles, culture, and more, it will behave much more like a company employee vs. an agent that only uses a generic LLM knowledge of the world. 

• Interactivity: Some agents will get an initial goal set and execute or fail without interacting with humans. Some are built to include the humans in the loop and seek guidance as needed, de-facto working as an additional team member. 

• How they engage with humans: Beyond the level of interactivity, agents will engage with humans in different ways - chat, voice or video conversation, different results format, and more. The style and tone of conversation can also differ significantly. 

• Guardrails and control: These will impact the degree of freedom and the ability of an agent to make mistakes. When building agents for regulated industries or tasks or where the cost of errors is high, one should apply much more control. 

• Multi-modal - in the broad sense of the term - the different modalities by which an agent can understand and interact with the world and user. 

• Interpretability - how much visibility (and impact) the human has into the different actions the agent plans or has taken. 

If you’ve read on to this point, you might wonder, “Is this agent thing really for me right now?” or how to go from theory to practice. 

There is science, but know-how and experimentation are also required to make agents more helpful than harmful. 

The bottom line is that there is no replacement for getting your hands and heads into the proverbial “agents pool” and starting to ideate and experiment. How? You might wonder.