A typical Agentic workflow may look like this;
The user provides a task E.g., "Write an excerpt about the solar system."
The agent understands the goal and breaks it down (using train of thought reasoning)
The agent selects a tool, such as a web search tool
Based on user instructions, the agent can retrieve the top 5 relevant sources from the web search
The agent evaluates the sources and then selects the most relevant one
The agent creates an initial draft based on the selected source
The agent reflects on the draft, checking accuracy, clarity, and more.
The agent delivers the final output, which is an accurate excerpt.
What powers an Agentic System?
Behind an Agentic system lies a large language model.
A large language model (LLM) is considered the brain or the reasoning engine of agentic systems. A Large Language Model (LLM) is capable of understanding instructions and generating responses, but when it is combined with:
It is transformed into something that can reason, plan, take action, and execute tasks.
This integration is what makes an agentic system capable of setting goals, following structured actions with minimal human input, unlike a typical chatbot that responds to user prompts and generates single responses.
Components of Agentic AI
Planning and Execution: Agentic AI systems do more than react to commands. This component allows systems to determine how to achieve a goal by breaking it into manageable tasks and deciding the steps required.
The AI system decides the appropriate tools, order of actions, and also evaluates potential obstacles. After planning, it executes, monitors progress, and updates as required.
Memory and State Tracking: This component allows AI systems to remember the past and track events in real time. By maintaining state, the system can use past experiences to make future decisions.
Advanced memory systems help the AI avoid repeating mistakes and learn from what worked or failed, improving performance over time.
Perception and Environmental Awareness: Agentic AI systems can interpret activities in their environment via visual, audio, and textual inputs. This enables them to identify trends, understand context, recognize patterns, and use real-time monitoring to make informed decisions instead of relying on predefined input.
Learning and Adaptation Engine: AI systems evaluate and observe the outcomes of actions and use the feedback to make more accurate future decisions. They review the result of previous actions, refine strategies, and adapt the approach where necessary.
Communication and Interaction: This component enables the AI system to collaborate with humans and other AI systems / external tools. It allows the agent to exchange information, determine, and plan actions to achieve and improve its objectives.
Decision-Making and Control System: This component is responsible for the agent's mechanism. It brings all components, including planning, memory, perception, communication, and learning, ensuring the AI behaves appropriately and performs actions based on the defined goal.
Benefits of Agentic AI
Parallelism: Parallelism in an agentic workflow means the agent performs multiple independent tasks simultaneously. For example, the agent can search different sources concurrently and evaluate them in parallel before deciding which information to use. This reduces the time required to complete tasks and improves efficiency.
Perform Multiple Tasks Effectively: Agentic AI can understand and break down complex goals into smaller tasks and decide how it wants to accomplish them to achieve a final result. The agent understands context and priorities and selects the right tools required for the tasks.
Modularity: Agentic AI enables complex goals to be broken down into independent modules responsible for specific functions, such as planning, reasoning, or evaluation.
Each module operates independently and can be reused across different workflows, making it easy to develop, maintain, and upgrade workflows without interrupting the entire system. This modular structure supports the development of flexible, scalable, and resilient systems.
Types of Agentic AI
There are different types of Agentic AI systems. They can be categorised based on autonomy, complexity, and applications. Here are some common types:
Autonomous Single-Agent System: This AI agent operates independently without other agents. It can plan, make decisions, and execute tasks within predefined boundaries. For example, a customer support agent may resolve queries, update CRM records, and send follow-up emails.
Multi-Agent Systems: Multi-agent systems consist of multiple agents that collaborate to achieve shared or individual goals. They are used in complex environments where a single agent system may not be able to handle task such as a resource management system or large-scale automation systems.
The different multi-agent systems include:
Vertical multi-agent systems: In this structure, the AI agent acts as a manager; delegates tasks to supporting agents, monitors progress, and ensures that the overall goal is achieved.
Horizontal multi-agent systems: This involves agents that work together as peers. Each agent specializes in a specific task and collaborates to solve problems and make decisions collectively.
Human-in-the-Loop Agentic Systems: Human-in-the-loop agentic systems involve AI agents that operate autonomously but are overseen by humans when making critical decisions.
The agent manages part of the workflow, makes decisions, and also seeks approvals when necessary. It is used in high-risk environments such as healthcare or finance.
Agentic AI VS Traditional AI Agents
AI Agents | Agentic AI Agents |
Often trained for Specific Tasks using predefined rules or models. | Performs tasks autonomously. It can plan and set goals. |
Operates with fixed instructions, which can limit operations. | Highly autonomous, meaning it can plan, execute, adapt, and handle complex tasks.
|
Cannot adapt to new situations without human input. Requires regular updates for new scenarios. | Can adapt to new situations without human intervention. |
Typically reactive, meaning it responds to prompts or instructions. | Proactive, meaning it can set sub-goals and execute independently. |
Good choice for single-step and routine tasks. | Designed for multiple-step tasks and processes that involve planning and coordination. |
Uses tools based on human instruction. | Can select tools to execute some tasks without human intervention. |