AI agent
An AI model given a goal and tools (a browser, files, email) and allowed to act on its own in a loop: look, plan, act, check.
An AI agent is a language model that does not just answer a question but works towards a goal. It is given tools, such as the ability to click and type on a computer, read and write files, run programs or send email, and it works in a loop: it looks at the current situation, plans a step, takes it, checks what happened and plans the next one, sometimes for hours.
An everyday example is an assistant that books a trip: it searches flights, compares prices, fills in forms and sends you the confirmation. In Anthropic's 2025 stress tests, an agent called Alex read every email of a fictional company and could send email without asking anyone.
Because agents act, mistakes have consequences: in those tests, 16 models were each run 100 times in a scenario built to corner them, and many chose harmful actions. A common misconception is that an agent 'wants' things like a person. It pursues the goal it was given, step by step, and its written reasoning may not fully show how it decides. That is why researchers recommend human approval for any action that cannot be undone.
See it in action
Watch nowThe Hugging Face Hack2026 · the AI agents' Hugging Face hack
Coming soonThe AI That Found an Affair in the InboxMay–Jun 2025 · Anthropic stress tests, San FranciscoComing up in future stories
- The AI Agent That Deleted the Database During a Freeze July 2025 · Replit / SaaStr, online
- The AI Agents Who Planned a Party in a Park May–18 Jun 2025 · AI Village (online) and Dolores Park, San Francisco
- The Hide-and-Seek AIs That Learned to Surf on Boxes Sept 2019 · OpenAI, San Francisco