AI agent

An AI agent is a program built around an LLM that decides what to do next, takes an action (calling a tool, writing code, clicking a button), observes the result, and loops until the goal is met or it gives up. The defining trait is autonomy in the loop: at each turn the model, not the developer, chooses the next step. The agent loop: 1. Sense — read the input, existing context, and available tools 2. Reason — decide the next action (or whether the task is complete) 3. Act — call a tool, write output, or ask a clarifying question 4. Observe — read the result, update context → repeat until goal met or step limit reached This is the structural difference from a chatbot. A chatbot answers your message and stops. An agent reads your message, plans, calls a search API, reads the results, calls a calendar API, drafts an email, asks you to confirm, and sends it. The "agentic" part is the model deciding the order and the stopping condition. The term is overloaded in 2026. Marketing pages call any LLM call with a tool an "agent." Anthropic's "Building effective agents" (December 19, 2024) draws the line most builders now use: workflows are "systems where LLMs and tools are orchestrated through predefined code paths," while agents are "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." Many production "agents" are workflows with one model call in the middle. That is often the right design, because a true agent is harder to debug and costs more per task, but knowing which one you are building changes how you design and evaluate it. Agents in production as of September 2026: Claude Code, and Cursor's Agent mode and Cloud Agents (read, write, and run code); Devin from Cognition (an autonomous coding agent); Replit Agent 4 (builds and deploys apps); and ChatGPT Work, which OpenAI launched on July 9, 2026 to act across apps and files and stay with a project for hours. ChatGPT Work replaced ChatGPT agent, which had in turn absorbed Operator, OpenAI's January 2025 browser agent. They share a structure: a tool-use loop, working memory, and explicit termination logic.