As the use of large language models has become ubiquitous, many technologists predict that LLM-based autonomous agents represent the future of generative artificial intelligence (GenAI).

While Agentic AI is still in its early stages, organizations need to anticipate its needs. Some enterprises have already deployed agentic workflows into their operational and customer-facing processes, but the full potential of autonomous agents is still being realized. In only a couple of years, it’s predicted that 33% of enterprise software applications will use agentic AI.

What are AI agents?

Traditional artificial intelligence (AI) needs instructions from a human to function. Chatbots like ChatGPT and AI assistants like Siri or Alexa require humans to provide well-defined instructions. For the most part, advanced large language models (LLMs) can perform straightforward tasks relatively accurately, but they do depend on human input.

Quickly emerging, the next generation of AI goes beyond simply prompting an LLM to summarize a document or draft a report. AI agents are expected to solve problems without step-by-step instructions. They’ll be able to work proactively and perform tasks autonomously.

In an interview with VentureBeat, Vijoy Pandey, Senior Vice President of Outshift by Cisco, said that AI agents are more online and work in real time compared to AI assistants. “They’re always listening and they’re pulling in data continuously.”

There are three common characteristics of AI agents:

Planning and reasoning — Once an agent is given a goal, it comes up with a plan and then acts on it to accomplish it. They use planning and reasoning to handle complex situations to make autonomous decisions, proactively adjust plans based on new information and feedback, and decompose big problems into small tasks.
Tools and services — AI agents have access to an array of different tools and online services, such as APIs, to complete their tasks. The agent can decide when and how different tools and services are used to provide accurate and relevant responses.
Memory — AI agents can learn from their experiences and adapt their behavior based on changing circumstances. With their long-term memory, they can examine their histories of interactions and data to create more personalized interactions, while using short-term memory to help with immediate tasks and processes.

Key concepts to know

The following essential concepts will help you better understand the underlying technology and capabilities of AI agents.

Autonomous agents

An autonomous agent operates independently to perform specific tasks, make decisions, and solve problems without needing constant human supervision. While autonomous agents are typically focused on predefined tasks, autonomous AI encompasses a broader range of capabilities.

Reinforcement learning

A machine learning technique that allows AI agents to learn from the environment and make decisions through feedback. It’s similar to trial and error to find the best way to accomplish a task.

Natural language processing

Natural language processing (NLP) enables AI agents to understand, interpret, and generate human languages. This allows for the AI system to effectively communicate with us.

Single-agent architectures

Just as it sounds, a single-agent architecture is one independently operating agent. They manage their own tasks and resources to complete their goals.

Multi-agent architectures

Multi-agent architectures have two or more agents and can use one or multiple language models. In this architecture, multiple agents coordinate, communicate and collaborate with each other.

Agent persona

It’s common for each unique AI agent to have its own persona, which defines the role or personality of the agent. The persona profile assigned to the AI agent also has any specific instructions that it may need related to its tools and services. Engineers can define and specify the persona of the agent within the system prompt, which is the predefined set of instructions that the AI system uses to generate replies and perform actions.

Breaking it down: How AI agents work

Think of an AI agent as a butler, designed to serve you with a specific persona that aligns to your needs and preferences. Over many interactions, the butler understands your habits and preferences and works proactively without needing direction. They’re experts at problem solving too.

For example, one of the butler’s jobs is meal prep. Every morning the butler serves you breakfast based on your preferences, routines, and contextual clues. They’ve noticed the time you eat and the types of foods you prefer. You don’t need to ask for breakfast each morning, it appears like clockwork. All the details have been accounted for.

In a multi-agent architecture scenario, there can be multiple types of AI agents with different personas assigned to different roles to achieve specific tasks. Let’s say you employ several workers to run your household, such as a chef, dog walker, butler, and nanny. This team works together and separately to ensure the household runs smoothly. It’s like how a multi-agent architecture functions.

Many AI agents have specific workflows they follow when they’re assigned to solve particular problems.

Planning

Once an AI agent receives initial direction from a user, it plans. It takes a complicated problem with many parts and decomposes it into smaller, manageable subtasks. The AI agents account for the user’s goals, constraints, as well as the tools and services it has available to create an initial plan. The planning process can involve various levels of complexity:

Simple planning: For run-of-the-mill routine tasks with a clear goal, the AI agent can create a plan with a limited number of steps.
Mult-step planning: For more intricate, interconnected tasks or a multi-phase project, the AI agent may go for a multi-step planning approach. It’ll break down the overarching tasks into a series of logical steps.
Re-drafting existing plans: AI agents are adaptive and flexible to emerging information. They’ll modify their plans as needed to accommodate changes.

Information gathering

For AI agents to work autonomously, they need information. This data can come from their specified environment, sensors, or user inputs. The agent processes this data to further understand the context and intent. During this information gathering phase, the agent starts to fill in the missing pieces it needs to accomplish its subtasks to ultimately achieve its overarching goal.

Decision making

Prior to acting, the AI agent reviews and analyzes all the information it gathers to evaluate its different options for completing each subtask. In this phase the AI agent is selecting the most effective strategy to complete each subtask.

Execution

After deciding on a strategy to tackle the work, the AI agent starts implementing the subtasks. As it completes each subtask, it inspects and analyzes the work, often looking for external feedback or assessing the results to ensure they meet the desired criteria.

Reflection

With every interaction, the AI agent continues to refine and learn. It updates its knowledge base so it can better align to future requests and improve its results. This continuous process of learning from feedback to improve an agent is known as iterative refinement.

Different types of AI agents

ReAct agents

The ReAct framework is an approach designed to combine reasoning and acting within LLMs to tackle tasks more effectively, reduce errors, and provide more reliable and interpretable solutions. A ReAct agent thinks through a problem, takes action to find new information, and then continue to refine its reasoning and actions until it produces an accurate solution.

Simple-reflex agents

As its name suggests, simple reflex agents are one of the most basic types of AI agents we have. Simple reflex agents act based on their immediate perception of the environment. When certain conditions are detected, it triggers an automatic response. Simple reflex agents don’t have a memory and instead rely on a set of conditions being met in the present to act.

A simple-reflex agent can be compared to a motion-sensor light. When motion is detected, the light turns on. If no motion is detected for a certain amount of time, the light turns off. The light doesn’t need to remember the last instance that the light came on.

Model-based agents

A model-based agent have an internal memory that helps them remember parts of the world that they can’t see at the moment. Just like a diorama is a small, detailed representation of a larger scene, the model in a model-based agent is a simplified, internal representation of the real world for the agent to use. This model allows the agent to understand and remember parts of the environment that it can’t perceive at any given moment.

Goal-based agents

Goal-based agents are built upon the foundation of model-based agents. In addition to having the same characteristics of a model-based agent, they also have a goal or set of goals they’re working to accomplish. Goal-based agents adapt and adjust in real time as they deem necessary to find the most efficient path to achieve their goals.

For example, a goal-based agent may monitor the water saturation of soil in a garden to ensure optimal moisture levels are kept. If necessary, it turns on the sprinklers and shuts them off when the appropriate level is reached.

Utility-based agents

Utility-based agents evaluate possible actions based on how well they align with predefined utility functions. These functions represent the agent’s preferences and goals. By choosing actions that maximize their utility, the agents ensure their behavior is consistent with rational decision making principes. Utility-based agents aim to maximize their own utility function.

Learning agents

Simple-reflex agents, model-based agents, goal-based agents, and utility-based agents can all be designed to become learning agents. Learning agents improve and adapt their performance through machine learning and data mining. They excel at operating in unknown task environments and learn through past experiences, sensory input, and feedback.

Uses Cases of AI agents

We are at the tipping point where AI agents will influence industries, job fields, and domains. They promise to change how we interact with technology and the ways we work.

Software engineering

AI agents have the potential for automating coding, testing, debugging, and documentation generation. Here are two examples frameworks that have shown promise in the application of autonomous agents.

ChatDev presents an end-to-end framework with multiple agent roles. These multiple agents coordinate with each other through natural language conversations to make the software development life cycle more efficient and cost-effective.

MetaGPT is a meta programming framework that is designed to streamline the software development process. The idea behind it is simple: it combines multiple roles, such as product manager and engineer, into a virtual team to oversee the code generation process. Within this framework, several LLMs act as specialized experts to tackle specific subtasks. This group of multiple AI agents collaborate to effectively complete code generation with minimal human oversight.

Customer support: Agentic AI chatbots

Many organizations have already deployed AI-powered chatbots to provide a personalized user experience, reduce costs, and give support to users at any time of the day. But chatbots powered by agentic AI are set to raise the bar.

Agentic AI chatbots will continuously learn in real-time from interactions with customers, to provide improved responses all to increase overall customer satisfaction. As needed, they’ll stay up to date with new information, remember customer preferences, and use tools in their workflows as needed.

An agentic AI future is near

As agentic AI evolves, the reasoning and planning behind LLMs will continue to improve. As enterprises, governments, and academia continue to invest resources into creating the next frontier in the AI field, the future of agentic AI is set to transform our daily lives.

At Outshift by Cisco, we are building an open, secure, and interoperable ecosystem for agentic AI. Join the community and follow along for updates!