OpenAI launches Dots, taking on Meta’s Muse in race for AI agents

OpenAI has launched Dots, a new class of “always-on” artificial intelligence agents designed to work on users’ behalf even after they have stepped away from the chat. The announcement, made at the company’s DevDay event in San Francisco, comes just after Meta introduced Muse earlier this month, its own personal AI agent designed to autonomously carry out tasks for users.

Sam Altman described Dots as “remarkably capable, always-on” agents, and said they represent “a whole new way to work with AI.” He also described them as an AI helper that always has one's back, adding that the company eventually sees users working with “a whole team of dots.”

Unlike a conventional chatbot, where a user gives an instruction and waits for a response, Dots are designed to take ownership of a task or project and continue making progress in the background. OpenAI says users can give a Dot a responsibility, connect the applications it needs and set rules around what it can and cannot do. The agent then works through the task and returns to the user with work to review or decisions that require their input.

OpenAI describes Dots as “an extension of you”, with the agents able to work across connected applications while continuing tasks in the background.

Each Dot has its own cloud computer, allowing it to browse the internet and use connected applications to complete tasks. According to OpenAI, Dots can connect to more than 4,000 apps, including workplace platforms such as Slack and Microsoft Teams. They can also use OpenAI’s Codex for coding-related tasks and ChatGPT Work for other work.

Dots can also draw on information stored in ChatGPT’s memory, allowing them to pick up on a user’s preferences and how they work. OpenAI, for instance, demonstrated how a Dot could check a user’s calendar, notice they were working through dinner and suggest ordering food. In another example, a Dot helped a user launch a website.

Users will also have some control over what these agents can do. They can choose which apps a Dot can access, set rules for its activities and decide when it must seek approval. OpenAI says additional checks apply to actions involving accounts or the sharing of information.

OpenAI follows Meta’s Muse

OpenAI’s announcement comes less than a month after Meta launched Muse on September 8. Meta’s personal AI agent is built to carry out tasks without requiring users to guide it through every step. The company says Muse can send emails, book travel, and work on longer projects. It operates through a dedicated virtual machine called Muse Secure VM, which houses the agent and user data.

Giving AI agents access to emails, calendars, and other applications, however, also raises security concerns. Meta has acknowledged these risks and detailed how it is working to protect Muse against threats such as prompt injection, where malicious instructions hidden in external content can influence an agent’s behaviour. OpenAI, too, has introduced permissions and approval requirements to restrict what Dots can do without a user’s consent.