Agentic AI: the next frontier in generative AI

AI is evolving at an unprecedented pace, reshaping how we interact with technology and the world around us. Generative AI (GenAI), a key innovation in this space, has demonstrated remarkable capabilities in creating human-like content, whether it’s drafting documents, designing visuals or generating videos. However, while GenAI revolutionizes content creation, the next big leap in AI—agentic AI—promises to redefine how we interact with the world around us. 

Let’s dive deeper into the evolution of AI through a real-world example: planning a vacation to visit four countries in Europe, say, Switzerland, Germany, Italy and England. Whether you like planning for vacations or not, we can all acknowledge that this endeavor can become complex quickly. Especially if you need precise planning day by day, book flights, hotels, rental cars, local tours, local transportation, etc., all within your budget.

With that backdrop, we will highlight the distinction between GenAI, AI agents and the emerging capabilities of agentic AI.

GenAI: creativity in action

GenAI excels at creating content based on prompts. Of course, you need to have robust models to generate human-like content. For instance, when asked, “Create a 10-day itinerary for Switzerland, Germany, Italy and England,” GenAI models such as ChatGPT or Google Gemini could draft a detailed itinerary. They might suggest landmarks to visit, such as the Swiss Alps, Berlin’s Brandenburg Gate, Rome’s Colosseum and London’s Big Ben. The output is often creative, well-structured and informative.

However, GenAI is limited to content generation in this context. While it can recommend destinations and activities, we still need to research transportation options, book accommodations and consider visa requirements. The model doesn’t take any action or adapt its recommendations based on real-time constraints, such as budget changes or availability.

AI agents: task-oriented execution

AI agents take things a step further by combining generative capabilities with task execution. Using the same vacation planning example, an AI agent could:

• Search for flights between the countries.
• Compare hotel options and book rooms.
• Purchase train tickets for inter-city travel.

An AI agent can handle predefined tasks using specific tools or integrations, such as travel booking websites. However, it often requires explicit instructions or operates within narrowly defined parameters. For example, if travel restrictions suddenly change or a preferred hotel is unavailable, the AI agent might not adapt dynamically without human intervention. Overall, these “simple or non-adaptive” AI agents operate with limited scope within the confines of its programming or environment.

Agentic AI: goal oriented autonomy

Agentic AI is the next phase of GenAI. While GenAI can generate content such as documents or images, Agentic AI is all about making AI serviceable for task-oriented executions and goal-oriented autonomy. Agentic AI incorporates the use of multiple AI agents to complete tasks autonomously, providing outcomes with very little human input for long, complicated tasks. Because of this autonomy, AI agents can become a conduit for activities we currently do manually, like planning a vacation abroad. For example, in the vacation planning scenario, agentic AI could:

• Understand the goal: Recognize the user’s preferences, such as budget constraints, interests (e.g., history, adventure, nature or cuisine) and travel style.
• Adapt dynamically: Respond to real-time factors including sudden flight cancellations, weather conditions or travel policy changes.
• Optimize outcomes: Recommend cost-effective routes, such as taking a train from Switzerland to Germany instead of flying, and adjust the itinerary accordingly.
• Act autonomously: Book flights, hotels and activities while ensuring all travel documents and insurance requirements are met.

Further, if a user’s budget unexpectedly decreases during the planning process, agentic AI could suggest budget-friendly alternatives without compromising on the overall experience. It could also monitor price drops for flights or accommodations and rebook accordingly, ensuring the trip remains within financial limits.

Overall, agentic AI addresses several key limitations of current AI systems with the following key capabilities:

• Autonomy: Unlike GenAI and basic AI agents, agentic AI can independently perform tasks and make decisions based on user-defined goals to deliver outcomes.
• Adaptability: It dynamically adjusts its approach as new information emerges, ensuring optimal results even in complex, changing scenarios.
• Efficiency: By integrating creativity, action, and reasoning, agentic AI reduces human effort, saving time and resources.

The anatomy of agentic AI

Note that at its core, agentic AI is all about goal-oriented problem solving. So, how exactly does an agentic model do all of this? Let’s dive into the anatomy of agentic AI.

To start, many agentic AI models are structured with multiple AI agents, each being an expert in its own space. Continuing our vacation example, AI agent 1 could be an expert in “flights.” AI agent 2 could be in hotels. And AI agent 3 could be in tourist attractions.

The core reason it’s beneficial for each AI agent to be its own expert is to greatly decrease the chances of hallucinations and inaccuracies. Ultimately, these AI agents will talk together to build a plan and then execute it autonomously with the parameters it was given. 

So let’s say we told our AI agents that the goal is to plan a safe vacation within our defined budget. 

The AI agents will then go through five steps, depicted in the following diagram, to achieve that goal:

Step 1: AI agents perceive or sense the external environment.

First they perceive or sense the external environment and various factors, such as all the many flight and hotel options that are available. In the vacation example, the agents would:

• Access airline and travel agency APIs to fetch flight schedules, prices and availability.
• Sense user preferences (e.g., travel dates, budget, preferred airlines, seats and meals) from direct inputs or historical data.

Step 2: AI agents reason these external factors.

Then, they reason these external factors by processing the information and making decisions. At a high level, this is powered by LLMs and could employ techniques such as chain of thought reasoning, reason-act (ReACT) or other techniques. This is when they will determine the best flight and hotel options for you on your behalf. For example:

• Compare flight options based on criteria like price, layovers, departure/arrival times and airline ratings.
• Identify the optimal trade-off between cost and convenience.
• Use predictive algorithms to determine if prices are likely to rise or drop.

Step 3: AI agents develop a plan.

Next they plan. The AI agents can then develop a sequence of actions to achieve the goal by confirming a checkpoint list of parameters like benchmarking the plan against your budget. For example, AI agents would:

• Confirm user preferences and budget constraints.
• Search for flights within defined parameters.
• Reserve the flight if availability matches the user’s preferences.

Step 4: AI agents act on the plan.

Then, the AI agents can act. This is where they will execute the plan through interactions with external systems or interfaces. This includes workflow orchestration, use of techniques such as RAG, etc. It will use APIs to reserve a flight and a hotel, process payment for reservations and even send you the itinerary. In our example, AI agents will:

• Use APIs to reserve a seat on the chosen flight.
• Process payment securely through a payment gateway.
• Send the user a booking confirmation and itinerary.

Step 5: AI agents learn and improve performance over time.

And finally, the AI agent will learn from all of these recent interactions. This is the most human-like part of the anatomy, where the AI agents learn from its mistakes or missed opportunities. Maybe it failed to register your preferred airline or hotel. Next time, it will add that to consideration to finetune its effectiveness even more. AI agents learn by:

• Analyzing user feedback or missed booking opportunities to refine future searches.
• Learning user preferences, such as preferred seating or loyalty programs, to tailor future recommendations.

Agentic AI frameworks: key takeaways

Agentic AI marks a transformative shift in the AI landscape. While GenAI has revolutionized creativity and AI agents excel in task execution, agentic AI combines the best of both worlds with autonomy and adaptability. From planning vacations to solving complex business challenges, it is poised to become an indispensable tool in the future of business and our daily lives. As we embrace this next wave of AI innovation, the potential for reshaping industries and enhancing human experiences is boundless.