AI Agent

An AI agent is a software system that uses a Large Language Model (LLM) to pursue a goal through multiple decisions and actions. The model can interpret the current state, plan or select a next step, use external capabilities, evaluate the result, and continue until it reaches a completion condition.

Unlike a chatbot that primarily generates conversational responses, an agent can change external state. It may search a knowledge base, call an API, edit a file, operate a user interface, or delegate work to another agent.

Core components commonly include:

• a goal and instructions supplied through the model's context;
• tool calling for accessing data and taking actions;
• an agent harness that runs the observe-decide-act loop and integrates tools;
• agent memory or external state for continuity;
• agent orchestration for routing, delegation, and control flow; and
• agent evaluation for measuring task success, cost, robustness, and safety.

Not every multi-step LLM application is fully autonomous. In an agentic workflow, application code may define a fixed sequence of steps. In a more autonomous agent, the model dynamically decides which steps and tools to use.

Reusable agent skills and agent scaffolding can provide specialized procedures and resources, while dynamic tool discovery keeps large tool catalogs out of the model's context until needed.

More complex systems may split work across a primary agent and a sub-agent, or transfer control through an agent handoff when another specialist is better suited to continue the task.

Some agents are reactive, responding only after a user message, while proactive AI initiates work from schedules, signals, or events. Event-driven AI is one architecture for triggering those workflows.

Greater autonomy increases flexibility but also expands the failure surface. Production agents require authorization boundaries, an explicit agent policy, step and cost limits, agent guardrails, human approval for consequential actions, protection against prompt injection attacks, and traces that record actions taken.

Long-horizon tasks usually need additional execution controls, such as an agent sandbox for isolation and a resumable agent run for recovering from interruptions without losing state.