AI systems that can take actions in the real world — not just generate text. Unlike the chatbots that first made headlines in 2022, agents can use tools: they can browse the web, send messages, write and run code, read and create files, and interact with databases and APIs. Crucially, the agent itself decides which tools to use and in what order, adapting its approach as it goes.

When LLM-powered chatbots like ChatGPT first arrived, users could have impressive conversations and get surprisingly useful answers, but the AI was confined to the chat window. It couldn't do anything outside of that text box. Agents changed this by giving language models access to tools — software that translates the model's text outputs into real-world actions. OpenAI's Operator, for example, can take control of a web browser to order groceries or make dinner reservations, while Anthropic's Claude Code can access files on your computer and spawn subagents to work on pieces of a task simultaneously.

For data reporters, agents open the door to automating multi-step workflows that are tedious but important. An agent could monitor a government website for new PDFs, download and name them consistently, extract tables into a dataset, and flag changes for a human to review. The trade-off is risk: when AI can take actions, mistakes have real consequences. Guardrails like approval checkpoints (see plan mode), least-privilege tool access, and defenses against prompt injection become critical. Related concepts include agent skills, subagents, Model Context Protocol, and the lethal trifecta.