Agentic Artificial Intelligence

Agentic artificial intelligence refers to an AI system that can pursue a goal through multiple steps, choosing and executing actions with limited human direction. In contrast, a conventional generative AI model primarily produces content, such as text, images or code, in response to supplied prompts. The distinction concerns the system's autonomy and architecture, not necessarily a separate kind of underlying model.

Core architecture

An agentic system commonly places a generative model inside a control loop that observes its environment, selects an action, evaluates the result and continues until it reaches a goal or stopping condition.

• It may decompose a goal into tasks and revise its plan when circumstances change.
• It can use external tools, such as software, databases or application interfaces, to obtain information or perform actions.
• Memory or stored state can help it retain relevant information across steps, although such memory is not essential to every agent.

How its operation differs

• A conventional model is mainly reactive: it maps a prompt and its context into generated content within a bounded interaction.
• An agentic system is goal-directed: it can determine intermediate steps needed to advance a specified goal.
• Conventional generation usually leaves execution to the user, while an agent may directly act through tools within its granted permissions.
• The difference is a continuum: systems vary in planning ability, duration, tool access, environmental reach and the amount of human oversight.

AI safety significance

Agency can convert an inaccurate or manipulated output into a sequence of consequential actions. Risk therefore depends not only on model capability, but also on autonomy, permissions, operating duration and the stakes of the environment.

• Errors may compound across multiple steps because later actions can rely on earlier mistaken outputs.
• Tool access can increase exposure to prompt injection, unintended disclosure and harmful execution.
• Safeguards include least-privilege access, sandboxing, human approval for high-impact actions, monitoring, audit logs and reliable stopping mechanisms.

How UPSC asks this