Generative AI Agents

Definition of AI Agents

To start off from a very basic definition, think of an AI agent as your personal assistant.  

When you ask your personal assistant to write an email, it works on it taking is as a goal. When you ask it to write an email, it will write an email for you, it does not play music for you or does not do something else. 

Hence the first thing that can be said about these agents is that they are intelligent and know what goals they are given and correspondingly what actions they need to take to get that work done. 

Purpose of AI Agents

Currently the AI agents are able to do a wide variety of tasks and that too extremely well. Some of those tasks are: 

• Browsing Web

• Write and Execute Code

• Book a meeting 

• Make a payment,

• Analyse and Query Data, etc. 

At this point of time where every day we see a new advancement in this field of Artificial Intelligence hence the current limitations on the list of things that an AI agent cannot do may become doable the very next day. 

Every new advancement that we see around Artificial Intelligence, Agents in particular, is taking us a step closer to AGI (Artificial General Intelligence). 

Types of AI Agents 

There are broadly 6 types of AI Agents

1. Simple Reflex Agents

It is more like hardcoding the agent’s behaviour. It basically works on condition-action rule meaning it acts after perceiving the current condition. They can neither plan their next move in advance nor can they learn and improve the reasoning by learning from past experiences.  

Although it is easy to implement and run and but is very inflexible to changes. Also, since this category of agents are not equipped with memory do not store any state, they have very limited or no intelligence of their own. 

One of the examples of this type of agents can be the rule-based chatbots which basically has a pre-planned set of responses for the queries of the user. 

2. Model-based Reflex Agents

It is similar to Simple Reflex Agents, but it also uses some intelligence of its own during the decision making.  This agent type work in a four-step process: 

• Sense: Here the agents gets to know the current state it is in before taking an action. 

• Model: In this step it makes a view for itself after sensing the current environment. 

• Reason: Based on the above created model it now decides how to act before actually taking any action. 

• Act: Here the agent actually carries out the action. 

An example of this type of agent is AWS Bedrock as it uses various foundational models for making decisions based on user prompts. 

Although these types of models are quick and better in decision making but they are computationally expensive. 

3. Goal-based Agents

They are different from the above two as they perceive information from their environment to achieve specific goals.  

They have three parameters that they take into consideration when they work: 

• Current state  

• The end goal to obtain 

• Set of actions needed to take to reach the goal 

These types of agents are very effective when deployed to attain a specific goal, but it may fail for a complex task. 

4. Utility-based Agents

These agents are quite advanced as they can assign utility score to different path it needs to take in scenarios when there is more than one possible path to complete a certain task. 

Consider a scenario when there is an agent designed to do research. But for a certain task it has both the options – search web or go through the vector store to complete a sub process. In this scenario this agent will be able to add utilities to these separate paths and then can decide which one to take to complete that particular task. 

The main advantage of these type of agents is that they can perform well in a wide variety of scenarios involving decision making. It also learns from previous experiences and accordingly adjust is decision making strategy. 

5. Learning agents

As the name suggests this type of agents can learn from the past interactions and with time improve the performance. The agents here learns from complex data patterns and may also receive feedback from humans in the loop to adapt accordingly. 

6. Hierarchical Agents

This is very similar to how things are executed in an organisation in a hierarchical manner.  

The agents in the lower-level hierarchy actually executes the tasks and their supervision is done by the agents present higher above them in the hierarchy. 

This agent type is very good when it comes to prioritizing different tasks by assigning the right set of tasks to the right agent. 

Use Cases of AI Agents

1. Personal Virtual Assistants

It is a very popular use case involving AI Agents. They can assist us in various tasks like reminding us of an important event of the day, planning our day, writing mails for us, planning meetings, etc.  

2. Healthcare

AI agents in healthcare industry can make a significant impact. From basic tasks of helping people with their medical queries to remarkable task of helping in drug discoveries are all possible through the use of AI Agents. 

There are already many pharmaceutical companies like Gilead Sciences and more which have already witnessed the potential of AI Agents.  When it used to take years for research now, they have accelerated the whole process making it all possible within months or even days.  

3. Finance

Financial intuitions can also leverage this power of AI agents to help them in fraud detection by learning of the transaction patterns from the previous data. 

Also, they can utilise agents to build customer friendly chatbots which will not only be fast with responses to user’s question but will also be more accurate. 

4. Education and Research

Here agents can help in research by making almost the entire World Wide Web accessible by the means of prompts in Natural Language. This has not only reduced the time required of manually going through the research papers but has also made best content accessible required for any research. 

Best Open Source Agents

1. AutoGPT

This is one of the first and highly capable open-source AI agent available. For a given goal, it first creates a set of sub tasks. It then goes through each of those sub-tasks to get the work done. In this process it many even divide the sub-task into further sub-tasks based on the complexity of the task. 

It is basically the Agent dividing a complex task into sub tasks and iterating through those sub tasks to get the work done.

2. AutoGen

It is another of those remarkable releases in the field of Autonomous Agents. It allows you to build a Multi-Agent Conversational Framework inside of your application for enhanced accuracy and better inference from the LLMs. 

For example, say you need to build an agent that can query your structured database. Here using AutoGen you can pass on the results of the initial query through various other agents in the middle before finally producing the output to the user. Each agent has a defined task and a role assigned to it. If at any step an agent finds the responses not satisfactory, it sends it back to the previous agent for re-evaluation. 

Current Limitations of AI Agents

1. Data Dependent

The backbone of any AI agent is the Large Language Model. Hence for an AI Agent the accuracy in response and intelligent behaviour of the overall agent is directly dependent on the richness of the data that the LLM was trained on.  

The Agents may become highly biased if the LLM in the backend is not trained on the right set of data. 

2. Limited Understanding of Context

Taking about the currently available open-source AI agents like AutoGPT and others, they have a short-term memory making it hard for them to hold on to the context in a longer conversation. 

3. Security Concerns

Since these AI Agents lacks common sense and ethical perspective, they can easily be made to work on goals with malicious intents.