Autonomous AI Agents

What counts as an autonomous agent

An autonomous agent takes a goal, plans steps and carries them out with tools such as a browser, a code runner or a file system, checking in with you only when needed. The tools here range from consumer products that run tasks in their own cloud computer, such as Manus and Genspark, to software engineering agents such as Devin and OpenHands, to building blocks such as Browser Use that let a model operate a real web browser.

Frontier labs now ship agent modes inside their main assistants too. Claude's paid plans include Claude Code and task handoff and scheduling, per Claude's pricing page. ChatGPT offers ChatGPT Work and Codex. Gemini 4 Argon added a tasks mode and connections to more than 40 apps, per Google.

Before you hand an agent real work:

• Start with a task you can check. Agents are best at well defined jobs with a clear finish line.
• Limit access. Give it only the accounts and files the task needs, and watch for prompt injection on web pages it reads.
• Review before anything irreversible, such as payments, emails or code merges.
• Check the safety research. METR measures how long a task frontier models can complete on their own, and lab safety frameworks set limits on autonomy.

For how frontier labs judge agent risk, see Preparedness Framework vs Responsible Scaling Policy. For which assistant suits long running work, see Claude Fable vs Opus vs Mythos.

Agents and the cost of running them

Agent work uses far more tokens than a single chat, because the model reads tool output, plans again and retries. On Claude plans, Claude Code shares one usage pool with your chats, and Claude Fable 5.1 can use up to half of the weekly limits on Max, per Claude's pricing page. On APIs, cache pricing matters most: a repeated system prompt billed at the cached rate can cost a small fraction of fresh input. Set a spending limit in your provider console before letting an agent run unattended, and log each step so you can see where time and money went.