Agent

An agent is a system that uses an LLM as a decision-making policy to run tools in a loop: observe context/state, choose an action (tool call or message), execute it, incorporate the result, and repeat until a stop condition is met. Unlike an AI workflow, the control flow is determined by the model at runtime rather than fixed at design time, with planning enabling the agent to decompose complex tasks into steps and adapt its approach as it executes.

Details

The agent loop is typically implemented and operated by an agent runtime. Some agents are filesystem agents, meaning their tool use is centered on a persistent filesystem (often in a sandbox).

An agent can also spawn subagents to handle delegated subtasks with their own context and tools, forming a multi-agent system. When the tool loop includes retrieval actions, the agent performs agentic RAG - dynamically deciding what to retrieve and whether to refine searches based on intermediate results. Agent memory allows agents to persist and recall information across sessions, while agent checkpointing captures execution state for resume, rollback, or forking.

Agents are often exposed to users through a conversational interface, but "agent" describes the backend behavior (a tool-using loop) rather than the UI; an agent can also run via an API, a background job, or an event-driven workflow without an interactive chat surface.

The combination of LLM-driven decisions and real-world tool access makes agents susceptible to a range of threats including prompt injection, tool misuse, data exfiltration, goal manipulation, user manipulation, and context poisoning.

Synonyms

AI agent, LLM agent, tool-using agent, autonomous agent