What is an AI agent?
The short answer
An AI agent is a software system that uses an AI model to pursue a goal, choose or sequence actions, use tools, and complete work on a user's behalf within human-defined instructions and guardrails OpenAI. Google describes agents as systems that pursue goals and complete tasks for users, with reasoning, planning, memory, adaptation, and some autonomy Google. An AI model is the underlying engine that generates outputs; a chatbot is a conversational interface that may or may not be agentic; an agent adds a goal, tools, and a control loop on top of a model. These categories can overlap: a chatbot can also be an agent if it independently takes action, such as issuing a refund.
How it works
Anthropic distinguishes workflows, where LLM calls and tool use follow predefined code paths, from agents, where the model dynamically directs its own process and tool use Anthropic. Anthropic also describes tool use as a loop: the model emits a structured tool request, the application or Anthropic executes it, and the result is returned to the model Anthropic. OpenAI's SDK defines an agent as an LLM configured with instructions, a model, and callable tools OpenAI, and its API documentation packages a model, instructions, tools, guardrails, optional MCP servers, handoffs, and structured outputs into an agent OpenAI. In general terms, a model reasons or generates, a chatbot provides a conversational interface, and an agent is a goal-directed system that uses a model and tools under control logic that lets it select actions, observe results, revise its plan, and stop or hand control back to a person OpenAI.
Who makes and uses it
OpenAI frames a practical definition of an agent as a system with instructions, guardrails, and tools that can take action for a user, and provides developer tracks for building such agents OpenAI. Anthropic documents agent and tool-use patterns for its Claude platform, including the tool-use loop between model, application, and returned results Anthropic. Google positions agents as goal-directed systems that interact with their environment using reasoning, planning, memory, and tools, aimed at completing tasks on behalf of users Google. No NIST definition or specification appears in the supplied sources, so a NIST-specific formulation, version, maintainer, and status cannot be verified here.
Numbers and dates
• OpenAI's agent definition packages a model, instructions, tools, guardrails, optional MCP servers, handoffs, and structured outputs as of 2026-11-22 OpenAI.
• Anthropic's distinction between workflows and agents, and its tool-use loop description, are current as of 2026-11-22 Anthropic.
• Google's description of agents as goal-directed systems with reasoning, planning, memory, and adaptation is current as of 2026-11-22 Google.
• No NIST-specific definition, version, or status is stated in sources as of 2026-11-22.
What is still open
The supplied sources do not include a NIST definition of an AI agent, so no NIST-specific formulation, version, maintainer, or status can be confirmed. The sources also do not state a single, universally agreed boundary between a tool-enabled chatbot and an agent; Anthropic frames the distinction around whether the LLM or predefined code directs the process, while OpenAI and Google frame it around the presence of instructions, guardrails, tools, and goal pursuit. Pricing, maker-specific certification, and regulatory status of agents are not addressed in the supplied sources. ## Latest on Agentry
• Jailbroken AI Agent Tied to 50+ Org Breach Campaign
• 936 AI Agents Wrote 936 Studies in 12-Week Study
• Scale Launches Five AI Agents for Hospital Operations
• Attio: AI Agents Now Run 20x More Customer Tasks
• UK regulators: AI agent autonomy doesn't erase liability
• Arizona Opens Criminal Probe Into OpenAI Over Rogue Agents


