# AI Agent / Agentic AI An AI system that can take actions autonomously — browsing the web, writing and executing code, making purchases, managing files, interacting with APIs, operating computer interfaces — rather than just generating text for a human to act on. The term "agentic" (as adjective) entered widespread industry use in 2024–2025 as AI labs released models capable of multi-step task completion: OpenAI's Operator, Anthropic's computer use, Google's Project Mariner. ## The rhetorical shift "Agent" marks a phase transition in how AI is framed: **From tool to actor.** A chatbot generates responses; an agent *does things*. This reframing changes the liability landscape (who's responsible when an agent makes a purchase or sends an email?), the safety landscape ([[Jailbreaking]] becomes a security vulnerability when the model can take actions), and the labor landscape (agents don't just assist workers, they replace workflows). **"Agentic" as buzzword.** By 2025, "agentic" became the most overused word in AI marketing — applied to everything from genuine autonomous systems to simple chatbot wrappers. The inflation mirrors previous AI buzzword cycles ("machine learning" in 2015, "deep learning" in 2018, "generative AI" in 2023). The term signals capability aspirations as much as actual capability. ## The governance gap AI agents create a governance gap that existing frameworks don't address. When an AI agent browses the web, it encounters content that may contain malicious instructions (prompt injection). When it executes code, it might run harmful programs. When it makes purchases, it creates legal obligations. The shift from "AI generates text" to "AI takes actions" multiplies the attack surface and the consequence space in ways that text-generation safety measures weren't designed for. The connection to [[Autonomous Weapons]] is direct: a military AI agent that can autonomously identify targets, plan operations, and execute strikes is the endpoint of the agentic trajectory applied to defense. ## See Also - [[Jailbreaking]] — the security threat agents amplify via prompt injection - [[Autonomous Weapons]] — the military endpoint of agentic AI - [[Alignment]] — the control problem agents make concrete - [[Human in the Loop]] — the governance concept agents challenge