In the case of Generative AI, artificial intelligence agents can only follow step-by-step instructions (called "prompts"),but "Agentic AI" agents can pursue goals of their own by planning ahead and adapting to any change of circumstances they encounter, often over an extended period of time.
How Agentic AI Works
Even though Agentic AI is powered by Large Language Models (LLMs) just as Generative AI is, minimal human oversight is required by Agentic AI because of the multi-step process agents use to make their own decisions:
- The agent begins by gathering the information it needs from whatever sources it can access (often referred to as Perception of the Environment). This usually includes the internet and database systems local to it but can also include external sensors.
- At the Analysis and Planning stage, the agent will interpret the data it has access to in order to generate possible courses of action. Using pattern recognition and rule-based reasoning, the agent will create a plan of action.
- Once the agent has a plan, it will begin the Action stage, using whatever tools it has available to it.
- In order to reach it's goal, the agent will evaluate whether each step of it's plan has achieved the intended result and make changes where necessary. This Learning and Optimization stage will mean that the agent's strategies will improve over time.
Real World Applications of Agentic AI
Here are some examples of the complex tasks that Agentic AI agents can automatically tackle:
- Customer Service: Agents can go beyond simply responding to customer queries and can troubleshoot issues by accessing and analysing customer information and previous support requests to suggest solutions. They can also interact with other systems, for example to provide an update on a customer's order.
- IT: Complex IT processes such as infrastructure management and software releases can be fully planned and executed by AI agents, who can often identify potential issues and take action before they cause problems.
- Research: Agents can gather and analyze data before offering their own hypotheses, often freeing human team members to focus on more difficult problems.
- Business: Across the world of business, Agentic AI agents are streamlining operations and reducing repetitive tasks for their human team members. For example in logistics, they can use their own initiative to identify bottlenecks or locate missing stock.
- Transport: Agentic AI is behind the complex, real-time decisions that autonomous vehicles have to make every second. This is an example of an agent using information from external sensors (such as GPS and cameras) and other services (such as traffic data) to constantly assess it's environment and plan appropriately.
Potential Disadvantages of Agentic AI
For the moment, integrating agentic AI agents into day-to-day processes is expensive and requires a degree of technical expertise. And because of the nature of Agentic AI, it can be difficult to trace everything that it has done and explain why. Without strong safeguards, they could therefore make poor decisions, or as we have seen in the news recently, act illegally in order to achieve it's goals. This kind of AI therefore raises important questions around data protection, regulatory compliance and liability.
