The Rise of AI Agents and Agentic AI
Two of the buzzwords on latest 2024 and on 2025 are those related to the use of agents and AI. Two of the terms used, sometimes used interchangeably, but with nuanced differences are AI Agents and Agentic AI. The key differences between them reside in their degree of autonomy; and learning, interaction and adaptation with the environment.
AI Agents
AI Agents are typically systems or programs designed to perform specific tasks autonomously within a predefined set of rules or controlled environments. They operate on behalf of a user or another system, making decisions based on their programming or training, and complete tasks without constant human oversight. Some of its key features are:
- Autonomy: AI agents can initiate actions independently without requiring explicit human commands, allowing for efficient task execution in various contexts.
- Goal-Oriented Behaviour: They possess the ability to set and pursue specific goals, making them effective in achieving designated objectives.
- Environmental Interaction: AI agents can interact with both physical and digital environments, responding to changes and adapting their actions accordingly.
- Learning and Adaptation: These agents can learn from experiences and adjust their behaviour in response to new data or changing circumstances, enhancing their effectiveness over time.
AI agents can include technologies like chatbots for customer service or virtual assistants for scheduling. They are limited by their design to tasks they've been instructed to handle:
- Siri, Google Assistant, and Alexa are virtual assistants that help users with tasks like managing schedules and answering queries but primarily operate based on predefined commands without autonomous decision-making capabilities.
Multi-Agents AI Systems
Any system that uses AI will need probably to provide its models (LLMs or another kind of models) with some kind of access to the real world. For example the possibility to call a search tool to get external and updated information and access to websites info; to call a coding tool in order to solve a task, etc. This capacity is sometimes referred as having agency or be agentic.
This new systems with AI capabilities has also enabled collaborative efforts among various specialized agents to solve complex problems more efficiently than a single agent could manage. GitHub Copilot, for example, combines multiple AI agents to assist developers in coding tasks, highlighting the benefits of collective intelligence in software development.
Devansh describes this kind of AI system as a system that...
...decompose a user query into a bunch of mini-steps, which are handled by different components. Instead of relying on a main LLM (or another AI) to answer a complex user query in one shot, (they*) would break down the query into simpler sub-routines that tackle problems better…
(It*) is a way to ensure your stuff works by making the AI take auditable steps that we can control and correct. This gives us better accuracy, fewer uncontrollable errors, and the ability to tack on more features.
(*) Devansh name this Agentic AI Systems
We can see these new systems as programs where an AI model output (i.e. an LLM) controls the workflow. The influence of the AI model (i.e. an LLM) input on the code workflow is the level of agency of the models in the system. HuggingFace states that in this definition, "agent" is not a discrete, 0 or 1 definition: instead the nivel of agency evolves on a continus spectrum, as we give more or less power to the AI model (i.e. LLM) on our workflow
The table below by HuggingFace illustrates how agency level when using an LLM varies across systems:
| Agency Level | Description | How that's called | Example Pattern |
|---|---|---|---|
| ☆☆☆ | LLM output has no impact on program flow | Simple processor | process_llm_output(llm_response) |
| ★☆☆ | LLM output determines basic control flow | Router | if llm_decision(): path_a() else: path_b() |
| ★★☆ | LLM output determines function execution | Tool call | run_function(llm_chosen_tool, llm_chosen_args) |
| ★★★ | LLM output controls iteration and program continuation | Multi-step Agent | while llm_should_continue(): execute_next_step() |
| ★★★ | One agentic workflow can start another agentic workflow | Multi-Agent | if llm_trigger(): execute_agent() |
Agentic AI
Agentic AI, on the other hand, refers to a more advanced form of AI that not only performs tasks but does so with a higher degree of autonomy, decision-making, and goal-oriented behaviour.
Agentic AI systems can adapt to new circumstances (they are able to operate in complex and dynamic environments), learn from interactions, and make decisions based on broader objectives rather than just following pre-set instructions. They possess the capability to plan, reason, and self-correct in pursuit of their goals, which might involve complex problem-solving or multi-step task execution. Agentic AI pushes the boundaries towards more human-like intelligence, where the AI can initiate actions, manage workflows, and even collaborate with other systems or humans. Some of its key characteristics are:
- Self-Initiation: Similar to AI agents, agentic AI can autonomously initiate actions, but with a greater emphasis on proactive behavior that seeks opportunities for improvement.
- Iterative Planning: Agentic AI systems can continuously refine their strategies based on past interactions and experiences, allowing for more effective decision-making processes.
- Proactive Behavior: They not only respond to current conditions but also anticipate future needs and take initiative to address them, which is critical in dynamic settings.
- Resilience: Agentic AI systems are designed to recover from setbacks and adapt to obstacles, enabling them to maintain functionality even in the face of challenges.
- Self-Efficacy: These systems develop confidence in their problem-solving and decision-making capabilities over time, fostering independence and improved performance
Examples of Agentic AI could be:
- Waymo’s Self-Driving Cars: These vehicles utilize multiple AI technologies to navigate autonomously. They collect data from sensors, make real-time driving decisions, and adapt to changing road conditions, showcasing advanced goal-directed behaviour.
