What Are Autonomous AI Agents?

A standard LLM interaction is a single turn: you give input, the model gives output. An autonomous agent is fundamentally different: you give it a goal, and it figures out how to achieve it through a sequence of actions that can span minutes, hours, or longer. The agent perceives its environment (through tool results, file contents, web searches), maintains a representation of its current progress, selects and executes actions, and updates its plan based on what it discovers. Real-world autonomous agents are already deployed: coding agents like Devin and Claude Code can take a GitHub issue and produce a working pull request by writing code, running tests, reading error messages, and iterating. Research agents can receive a question, search the web, read papers, synthesize findings, and produce a cited report. Computer-use agents can navigate a desktop UI, fill out forms, and complete multi-step workflows. What makes an agent 'autonomous' is the degree to which it handles unexpected situations without human intervention: a brittle agent needs a human at every decision point; a robust autonomous agent handles ambiguity, errors, and surprises gracefully and knows when to escalate versus when to proceed. The degree of autonomy appropriate for any application is determined by the reversibility of the agent's actions, the cost of errors, and the reliability of the agent on the task class in question.