Published on 00/00/0000
Last updated on 00/00/0000
Published on 00/00/0000
Last updated on 00/00/0000
Share
Share
INSIGHTS
7 min read
Share
The interest in artificial intelligence (AI) has skyrocketed, thanks to the rise of generative artificial intelligence (GenAI) in consumer-friendly applications such as ChatGPT.
However, GenAI in its standard state is reactive, not proactive. Give it a prompt or send it a message, and it provides you with a response. In this form, it is a useful tool, but still one that cannot take action on its own. This is where AI agents come in.
An AI agent is an AI-based entity that can take actions on its own based on its environment. This can be a physical entity such as a robot or drone, or it can be a software-based entity that operates virtually. Think of AI agents as the next step in automation.
A large variety of AI agents specialize in performing different tasks. This specialization is one of the strengths of AI agents. A general GenAI application might be good enough for many tasks. However, specialization leads to enhanced productivity and capabilities when it comes to business. As enterprises consider leveraging AI agent systems to bring business benefits, they must be familiar with the different types of AI agents, with examples that demonstrate the strengths and weaknesses of each.
All AI agents have ways to sense the current environment, decide what to do, and then take action on the environment. AI agents are primarily categorized by how they decide what to do.
Model-based agents maintain an internal representation of the environment, which is updated based on gathered data. From a human perspective, this would be considered “memory.” Based on this memory, a model-based agent then chooses an action to follow, typically according to predefined rules and conditions.
The main advantage of a model-based AI agent is that it is predictable and efficient. Based on the predefined rules and conditions, decision-making is performed quickly. Additionally, since it is equipped with its internal model of the environment, the model-based agent has sufficient understanding and knowledge to make informed choices, as well as adapt to the current environment state.
The disadvantage of model-based agents is that creating an accurate, useful model of the environment can be challenging and computationally expensive. If the model is not built or utilized well, the agent will make incorrect decisions. Additionally, if the model becomes outdated, the agent’s effectiveness will decrease.
One example of a model-based agent is an automated customer service system. These systems typically contain a model of some information about the customer and the current tools available to it. Based on this model, the system chooses an appropriate response and action.
In some cases, simply modeling the environment is not enough. This is where goal-based agents come in.
Building on the internal memory of model-based agents, goal-based agents attempt to predict the effects of their actions before deciding which action to take. Humans supply the agent with a set of goals, which is essentially a desired environment state. By combining this predictive ability with the given goal, a goal-based agent can then decide which action to take to achieve these goals.
The advantage of goal-based agents is that behavior does not need to be tied to specific conditions or rules. The agent can autonomously take actions towards the predefined goals. This can lead to straightforward, efficient agents. When the goal can be clearly defined and measured, the goal-based agent can perform effectively.
However, in chaotic environments or when complex goals require a moderated, balanced approach, the goal-based agent can struggle with choosing an appropriate action.
An example of a goal-based agent is a vacuuming robot. The goal is clear: Cover as much floor area in the shortest amount of time possible. The robot predicts the effects of its actions, such as moving to different areas, avoiding obstacles, and optimizing its path to ensure maximum coverage quickly. Because the goal is concrete and measurable, the goal-based agent can determine the best approach to performing its task effectively.
Utility-based agents take goal-based agents to the next level. While still maintaining the predictive ability of “what will happen if I do a given action,” utility-based agents measure the value of this action based on maximizing the value of a given function. In essence, it is a goal-based agent where the goal is to have the highest “utility.”
The advantage of this approach is that the utility function can be defined to incorporate many different priorities, each with varying weights and importance. This allows the utility-based agent to operate in complex and changing environments successfully.
However, the downside of utility-based agents is that this utility function becomes all-encompassing. If incorrectly defined, the agent will follow it unquestioningly, leading to possibly undesirable actions.
An example of a utility-based agent is a recommendation system, famously used by media streaming services such as Spotify and Netflix. Based on the agent’s knowledge of your listening and viewing history, these systems attempt to find the next piece of content you are most likely to view and enjoy, thus seeking to maximize your enjoyment of the service.
An AI learning agent takes the above types of agents and adds the ability to evaluate the success of its actions. Whereas the other agents have their decision-making processes defined ahead of time, a learning agent can change the means of deciding which actions to take. Through feedback from the environment or direct human input, the agent attempts to improve its own efficacy.
Naturally, a learning agent's advantage is its ability to improve and adapt autonomously. This can decrease the need for manual intervention and maintenance costs.
The downside is that these agents can have expensive upfront costs. Additionally, a system that learns on its own can learn incorrectly, either intentionally by adversaries or simply by becoming misguided.
An example of this is an autonomous driving system. It learns from real-time data and feedback to improve its ability to navigate and make driving decisions. However, the use of autonomous driving systems comes with the warning that a human should always be ready to take over control of the vehicle. Learning agents can be powerful, but the possibility of error or misguided learning is still a real one.
When deciding which agent type to use, examine the complexity of your business problem and the potential value that AI agents can bring. Based on those considerations, try to find an alignment between implementation cost, agent value, and your available budget.
An AI agent’s implementation and maintenance costs are also tied to how easily it integrates with your existing infrastructure and workflows. Naturally, tighter compatibility will lead to smoother integration, reducing your implementation effort.
Next, evaluate data privacy and security concerns. While it may be simpler to spin up an AI agent based on an external system, this approach may not satisfy regulatory requirements for high-compliance industries such as healthcare or finance.
Finally, consider multiple-agent systems for more complex tasks with higher requirements. By linking multiple agents, complex tasks that require more planning and execution can be handled.
For example, in a complex transportation system that requires oversight of transport times, logistics, and safety, multiple AI agents can work together to handle such tasks autonomously and safely.
Specialization is key when considering the use of autonomous AI agents. You can achieve maximum value by matching the appropriate agent type with the goal you’re looking to achieve. Start small, experiment, and iterate. With the growth of GenAI and AI agents, there’s never been a better time to begin your journey towards AI automation.
For more reading on AI, check out four ways AI personalization can improve your customer experience strategy.
Get emerging insights on innovative technology straight to your inbox.
GenAI is full of exciting opportunities, but there are significant obstacles to overcome to fulfill AI’s full potential. Learn what those are and how to prepare.
The Shift is Outshift’s exclusive newsletter.
The latest news and updates on cloud native modern applications, application security, generative AI, quantum computing, and other groundbreaking innovations shaping the future of technology.