What Are AI Agents, and How to Implement Them

Photo by Igor Omilaev on Unsplash

Some experts believe that AI agents are just a passing fad, while others dream of AI-run banks and artificial CEOs delegating work to AI agents and humans alike. Our artificial intelligence consulting company believes that this technology has great potential, but there are limitations that firms considering implementing AI agents need to be aware of.

So, what are AI agents? What can they do in theory and in practice? And which challenges can you expect during implementation and adoption?

What is an AI agent?

The global autonomous AI and autonomous agent market is growing fast. It was worth $4.8 billion in 2023 and is expected to skyrocket to $28.5 billion by 2028, growing at a CAGR of 43%.

So, what are AI agents and how do they operate?

AI agent definition

An AI agent is a sophisticated software program designed to interact with its environment and perform specific tasks by mimicking human intelligence. It can analyze data, make decisions, take actions autonomously, learn from past experiences, and adjust its behavior over time.

AI agents can be physical, like a smart thermostat, or software-based, such as a virtual personal assistant living inside an enterprise resource planning (ERP) system.

This technology became popular after the emergence of OpenAI’s ChatGPT, as this large language model makes it easier for computer programs to understand human language and interact with people. Generative AI can also handle large amounts of data and analyze it faster than classic AI. But it’s worth noting here that not every AI agent has to be based on Gen AI. It can function just fine with pure artificial intelligence, depending on its field and responsibilities.

Check out our recent article on Gen AI vs. artificial intelligence to understand the difference between these technologies.

How do AI agents differ from other software programs?

What differentiates AI agents from AI-powered chatbots and robotic process automation software is that AI agents are rational. They make decisions based on the data they receive and their interaction with the environment. They can react differently to the same situation, depending on the input and the environment. Users don’t have to constantly send prompts, like in ChatGPT, to force an AI agent to act. It will make decisions and take intermediate steps independently to reach the final goal.

Chatbots are built to interact with humans, while AI agents are supposed to work autonomously and only contact end users to deliver the final response or escalate a task if they don’t have enough information to see it to completion.

Key characteristics of AI agents

You can distinguish AI agents from other software through the following characteristics:

• Autonomy. AI agents can make decisions and act independently without human intervention.
• Reactive and proactive. When interacting with the environment, AI agents can be reactive, responding to environmental stimuli, or proactive, taking initiative while moving towards their end goal. They can operate in a static environment with fixed rules or in an ever-changing dynamic environment, where agents need to constantly learn and change.
• Ability to learn and adapt. They rely on machine learning models to improve performance over time, learn from experiences, and adapt to environmental changes.
• Goal-oriented behavior. AI agents are programmed to achieve specific objectives. So, they have the tools to make decisions, prioritize tasks, and change their course of action to reach the end goal.

Types of AI agents

AI agents come in different types. Below, you can find each type explained with an example.

Simple-reflex agents

Their response to stimuli is based on predefined rules without considering possible consequences or past events. If their environment changes, simple-reflex agents can’t adapt unless someone changes the rules that guide their behavior. You can say these AI agents have “limited intelligence,” but they are effective in static environments, executing tasks that depend solely on the current perception.

One example of a simple-reflex AI agent is a basic smart thermostat without self-learning capabilities. It operates autonomously, using a temperature sensor to stabilize room temperature. When the temperature falls below the threshold, the AI agent activates the heater. When it exceeds the set point, the agent turns the heater off. Its behavior is purely reactive, relying on if-then rules.

Model-based reflex agents

They gather information about the world and maintain a model that helps them infer data that is not explicitly presented. Unlike the previous AI agent type, model-based reflex agents respond to stimuli by considering experiences and probable outcomes. These AI agents can also function in dynamic environments, as they can update their intrinsic model according to changes in the world.

You can take an intelligent vacuum cleaner as an example of a model-based reflex AI agent. It navigates and cleans a house by using an internal model of its environment. This agent continuously receives input from sensors that detect obstacles, dirt, and floor types. Based on this data and its internal map, it determines the most efficient cleaning strategy, like adjusting suction power on different surfaces.

Goal-based agents

These AI agents possess superior reasoning skills compared to the previous two types. They can investigate several paths leading to the end goal and choose the most efficient one. Goal-based AI agents can generate a list of sub-goals that they will follow. They will take action only if it moves them closer to the final goal.

An AI chess player exemplifies a goal-based AI agent by striving to win the game through strategic planning. It uses a comprehensive model of the chessboard and follows the rules of the game to evaluate possible moves. This AI agent sets a primary goal of achieving a checkmate against the opponent while minimizing the risk to its own pieces. It tries to anticipate the opponent’s moves and assess the outcomes of various strategies.

Utility-based agents

Utility-based AI agents evaluate different approaches and select the best-suited path toward the end goal based on utility (preference). If you use a utility-based agent to plan your trip to some destination, you can specify your preference to get there as soon as possible or as cheaply as possible, and your AI agent will select the corresponding course of action.

Let’s assume that a smart thermostat operates in a utility-based mode where the preference is energy efficiency. This thermostat uses sensors to monitor the current temperature and humidity levels, but it will also consider energy costs. It will lower the temperature slightly during peak energy prices, even if the house is fully occupied.

Learning AI agents

These AI agents can learn from their surroundings and experience. They begin with a limited knowledge set and expand their horizons with every completed task. These agents are a perfect fit for dynamic environments, as they can adapt automatically without the need for manual rule reprogramming.

A personalized recommendation agent is an example of a learning AI agent, as it continuously improves its suggestions by learning from user behavior and preferences. Initially, it recommends content based on its general popularity. As the user interacts with the platform, watching shows and movies, rating content, and browsing different genres, the AI agent collects and analyzes this data to understand the user’s preferences and make personalized content suggestions.

Hierarchical agents

These AI agents are organized in a hierarchy where higher-level agents decompose a task into subtasks and distribute them to lower-level agents. Lower-level AI agents execute these subtasks and report the results to their superiors.

A self-driving car can be seen as an example of hierarchical AI agents. A high-level planning agent determines the optimal route to the destination, considering traffic conditions and legal restrictions. It delegates tasks to mid-level agents responsible for specific behaviors like highway driving, city navigation, and parking. These, in turn, instruct low-level control agents to perform precise maneuvers such as steering, accelerating, and braking.

What are the components of AI agents?

A typical AI agent has six components:

Image by the author

• An environment is the domain where AI agents operate. It can be either physical or digital space.
• Sensors enable AI agents to collect information from the environment. This can be a simplistic data collection device, like a temperature sensor, or an interface to gather textual or audio input.
• Actuators convert an AI agent’s output into physical actions in the real world. Examples of such actuators are a robotic arm that moves stock and software that creates files or sends emails.
• A decision-making mechanism is the equivalent of a brain. This unit processes all the information coming from sensors and enables AI agents to decide what action they will take next. This mechanism can be a rule-based system, neural networks, or any other system that can cope with decision making.
• A learning system allows AI agents to learn from past interactions with the environment and improve performance over time. This component uses techniques, such as supervised, unsupervised, and reinforcement learning.
• Knowledge base is an accumulation of rules and facts that AI agents use to make decisions. As agents learn, the knowledge base expands over time.

How AI agents work

After learning what AI agents are, let’s see how they function. For the sake of simplicity, we can break down AI agents’ workflow into five high-level steps.

1. Goal initialization. An AI agent receives a request from the user. For example, the user can ask a customer service AI agent to analyze a client’s support ticket.
2. Creating a subtask list. Based on the request above, the AI agent will generate a prioritized list of subtasks that it needs to execute to reach the end goal.
3. Decision making. For each subtask, the AI agent collects data from the environment through various interfaces and sensors. This could involve searching the Internet, interacting with other software, and using sensors to measure various parameters, such as temperature and pressure. AI agents process the collected data using advanced algorithms and involve their knowledge base to make informed decisions.
4. Taking action. Based on its decisions, the AI agent performs specific actions, such as sending notifications, making recommendations, etc.
5. Learning to adapt. The AI agent continuously learns from its actions and outcomes to improve its performance over time, adapting to the changing environment.

Image by the author

Zooming in on AI agents’ real-world business applications

Recent research by Accenture shows that 96% of executives are confident that AI agent ecosystems hold promises for their organizations in the coming three years. So, which areas can they benefit the most? Let’s explore.

AI agents in healthcare

There are many benefits for AI in healthcare. What do AI agents mean for the field?

When it comes to patient treatment, AI agents can analyze patients’ symptoms and medical history, direct them for a scan if needed, analyze the resulting medical images, and report to the assigned doctor with a suggested diagnosis and a personalized treatment plan.

For instance, Healthcare Advisor AI from Tars can interact with patients, assess their symptoms, offer personalized recommendations, send educational materials and appointment reminders, and more while helping patients navigate their health conditions.

AI agents can also optimize hospital workflows by predicting patient admissions and improving resource allocation. For instance, an AI agent can predict that ten patients will be admitted this evening, and it will immediately adjust doctors’ schedules to accommodate this assumption.

AI agents in customer service

These AI agents can tend to customers 24/7, giving efficient and personalized support. They can answer questions, process refunds, troubleshoot technical issues, and much more. So, an AI agent receives a request, accesses the company’s database, the Internet, or any other source, processes this information, and takes action or escalates the matter to a human operator. AI agents learn from each interaction and can provide better responses over time.

The generative AI agent startup Beam offers ready-made, pre-trained AI agents for customer service, along with an opportunity for companies to customize their own tools. If you opt for off-the-shelf AI agents, you can deploy them in seconds. They are GDPR compliant and can run locally or in the cloud.

AI agents in manufacturing

AI agents can take over predictive maintenance, product quality assurance, and factory floor monitoring.

• Predictive maintenance. Factory managers can task AI agents with monitoring all the equipment, looking for anomalies. These agents will constantly check device sensor data and performance metrics to spot any deterioration in the early stages.
• Quality assurance. AI agents rely on sensor data and visual inspections to spot product defects, even at microscopic levels.
• Factory floor monitoring. AI agents can use computer vision to collect and analyze data about the manufacturing processes and report any safety hazards, such as leakage and violations of employee safety protocols. MotionMind AI for industrial safety is one example of AI agents monitoring manufacturing facilities.

Unlike regular artificial intelligence models that can also detect anomalies, AI agents can react. They have actuators, which allow them to adjust device parameters, activate an alarm system, stop a conveyor belt, and more.

AI agents in finance

AI agents acting as financial advisors can communicate with customers, analyze market trends, consider customer preferences, and suggest assets for customized investment portfolios. AI agents can also autonomously manage these portfolios and protect them against market fluctuations.

Another exciting example of AI agents in finance is fraud detection. These agents can study cybercrime patterns and monitor user activities and financial transactions in real time to detect fraudulent behavior. They can flag suspicious transactions and terminate them automatically. AI agents can also learn to scan financial software for security vulnerabilities and fix or report them before malicious actors take advantage.

If you are interested in this sector, you can learn more about Gen AI in finance and banking on our blog.

AI agents in transportation and logistics

From autonomous vehicles to logistics and delivery route optimization, AI agents are thriving in the transportation domain. Self-driving cars rely on AI agents to navigate, avoid collisions, abide by traffic rules, and continuously adapt to their environment. In addition to driving, AI agents can monitor the vehicle for anomalies and alert its manager.

When it comes to logistics and delivery, AI agents can analyze road conditions, weather, and vehicle performance and combine this with delivery schedules for optimal route planning. They can process real-time traffic data, autonomously redirect vehicles to less congested routes on the fly, and adjust delivery schedules accordingly. If you need more inspiration, check out our blog post on AI use cases in supply chain.

If you live in Phoenix or San Francisco, you can already try the autonomous ride-hailing service from Waymo. You won’t see a human driver in the cab. These cars are fully operated by AI agents. They rely on a comprehensive set of sensors to navigate during the day and at night. Each autonomous car is trained on over 20 billion miles of simulation and has around 20 million miles of real-road driving experience. Waymo claims their cars minimize accidents and injuries in the areas where they operate.

Getting real about AI agents with ITRex

So, can companies already build and successfully adopt AI agents in their current state? Or is the technology still in its infancy and can only offer limited benefits? ITRex CTO, Kirill Stashevsky, explains.

Theoretically, it’s possible to implement any of the AI agents described in this article. However, in practice, it will be a challenge to achieve accurate and predictable results at scale. For instance, it takes a few weeks to build a workable prototype of a teaching assistant AI agent based on a GPT interface that will produce results with 60% accuracy, but it will take at least ten months to take the accuracy level up to 90%.

And don’t expect a smooth, gradual improvement. AI models have a black box architecture, meaning it’s often hard to tell what exactly is wrong and which algorithm is skewing the results. You can raise an agent’s accuracy level to 70%, and the next day it drops to 50%, and you are left guessing why.

Generally speaking, AI agents in their current state are best suited for applications where slight output variations are acceptable. Otherwise, we strongly recommend implementing human verification mechanisms.

Limitations of AI agents

Today, AI agents have two key limitations:

• Accuracy and predictability (or lack thereof). As explained above, AI models are black boxes, and it’s a challenge to understand what exactly is going wrong and which component is failing. And if an AI agent relies on a large language model, it can hallucinate, generating a plausible answer that is not factually correct, which further exacerbates the situation.

One bank wanted to use Gen AI agents in customer support but was worried about incorrect answers. They implemented an architecture that includes fact checking every text produced by the AI agents. If the text is correct, it’s released to the client. If not, the agent refrains from answering and puts the client through to a human operator.

• Scalability. McKinsey discovered that industry leaders report promising results from using AI agents in controlled environments but find it hard to scale these operations. This can be traced back to accuracy issues, but there are other factors involved. Using AI at scale necessitates rewiring some organizational functions, cleaning the data, implementing data management frameworks, and ensuring the data is compliant and free of bias, etc.

Other limitations include:

• High initial investment. Developing AI solutions can be expensive. You will need to invest in software, hardware, data, and talent. For more information, check out our recent articles on the expenses associated with using artificial intelligence and how much it costs to implement generative AI.
• Integration with the existing systems. To use AI, you need to integrate it into your business workflows. This task can be rather complicated if you use legacy systems. You will also need to aggregate data from disparate sources into one unified system, which AI agents can access.
• Ethical and legal concerns. It’s difficult to navigate the legal landscape around AI, especially in heavily regulated sectors, such as healthcare and finance. Also, AI agents can produce biased and inaccurate results. Who is responsible when AI models cause harm? And who is liable for the incorrect decisions made by AI agents?

Opt for ITRex as your AI partner

AI agents can serve as a powerful tool to automate tedious tasks and streamline operations. However, in their current state, be conscious when entrusting them to perform critical tasks because, ironically, while being used to minimize human error, AI agents themselves can make mistakes. Even AI models built by giants like Google. Just recently, Google’s large language model Gemini was generating historically impossible images, such as Nazi soldiers with Asian features.

At ITRex, we can help you build AI agents based on publicly available platforms or develop them from scratch. We understand which obstacles to expect during implementation and how to move forward. Our diverse team has experience in machine learning, large language models, and generative AI development, as well as in IoT and embedded systems. We have dedicated data consultants who can help you with data strategy and management.

Our developers can build an AI agent pilot version that operates in a controlled environment in just a few months. But achieving stable and predictable results with such a complicated system takes around 10 months of testing and experimentation, depending on your desired level of accuracy and stability.

We also have an AI PoC service that can help you speed-test and validate your AI concept.

If you want to start by implementing regular AI, check out our guide on how to create an AI system and a post on how to implement AI in your organization (with a link to a free eBook inside).

Want to experiment with AI agents? Drop us a line! We will allocate you a diverse team with experience in large language models, firmware, and sensor integration. We can build both software-only AI agents and robotic agents that can take physical actions.

Originally published at https://itrexgroup.com on July 24, 2024.