AI Agent Levels
Agents and Assistants
AI agent or AI assistant is a generic term for a system that works with an LLM. It takes the user input. Then it builds the output using the LLM and the tools.
ChatGPT, Gemini, etc.
AI agents are the most commonly used name for systems that interact with LLMs.
Agent and assistant are used as the same thing. A subtle difference exists though. To see it, think of the real world meaning of the two terms.
Think of an assistant as a personal assistant with limited powers.
An agent/agency has more independent powers. We just ask an agent to do something for us.
Assistant Messages
All messages from an LLM are grouped under the role 'assistant'. This covers CoT reasoning, tool calls and results, and sub-agent calls.
Agentic AI
These are systems that take a goal from the user. Then they do a series of actions to reach that goal.
Reasoning -> Planning -> Execution
All agentic AI systems are AI agents, but not all AI agents are agentic AI systems.
An AI agent can be any system that works with an LLM.
An AI agent is reactive. Agentic AI is proactive.
Example - say you want to book a holiday. The agent first reasons about the flight booking. Then it plans the steps to book the flight and the hotels. Then it does them with no human help.
Agent autonomous levels
The below picture describes autonomy levels from lowest to highest.
