New AI agents that can do tasks for you
Anthropic, Google DeepMind and OpenAI have each just unveiled AI agents with varying degrees of capability to control a person’s computer just like a human would - moving the mouse, clicking buttons, typing text and navigating through different applications to complete tasks assigned by the user.
Anthropic's Claude
Anthropic was the first to announce computer control capabilities. It uses visual navigation through screenshot analysis and is available to developers building on Anthropic's models. It can operate across any desktop application.
Google's Project Mariner
Built on Gemini 2 language model, Project Mariner is currently restricted to Chrome browser and access is limited to trusted testers.
OpenAI's Operator
OpenAI’s AI agent is available to ChatGPT premium subscribers (US$200/month). It is primarily focused on browser-based tasks, uses chain-of-thought reasoning and requests user confirmation for final actions as a built-in safety feature.
AI agents signal a shift in how we might interact with computers and manage work in the near future. While current popular AI tools can generate content or provide suggestions, these new agents can actually execute tasks - from data entry, document processing and filing to scheduling and web research - potentially automating entire workflows.
They will drive the transformation of some roles from task execution to strategic implementation and oversight of AI agents, prompt engineering and critical judgment for reviewing AI actions. However, limitations and risks remain, including data security implications and the need for clear protocols about which tasks are suitable to be delegated to AI.