Agents are autonomous AI systems that can perform complex tasks, make decisions, and interact with their environment without continuous human intervention. These agents leverage the advanced capabilities of LLMs, such as natural language understanding, reasoning, and generation, to operate independently and achieve specific goals. There are three key characteristics that distinguish agents from single-shot LLM responses:
  1. Iteration: Agents engage in multi-step processes, continuously refining their actions based on feedback and new information. Unlike single-shot responses, which provide a one-time output based on a given prompt, agents iterate on their own outputs, allowing for more dynamic and adaptive behavior.
  2. Tool use: Agents can interact with external tools and systems to gather information, perform computations, or execute actions. This ability to use tools enables agents to extend their capabilities beyond the knowledge and skills inherent in the LLM itself. By integrating tool use into their decision-making process, agents can solve more complex problems and adapt to a wider range of scenarios.
  3. Planning and workflow: Agents are designed to break down complex tasks into smaller, manageable steps and create structured workflows to accomplish their goals. They can prioritize subtasks, make decisions based on intermediate results, and adjust their plans as needed. This planning capability allows agents to handle multi-faceted problems that require a sequence of coordinated actions.
LLM agents maintain an understanding of the ongoing context and use this information to guide their actions. They actively work towards achieving specific objectives or goals by selecting appropriate strategies, adapting to challenges, and learning from their experiences. The autonomous and goal-oriented nature of LLM agents enables them to operate effectively in a variety of domains and scenarios, making them well-suited for agentic workflows.