OpenAI’s Dot agent is enterprise software that can also order your dinner

OpenAI launched an agent platform named Dots, which functions as enterprise software capable of executing tasks and accessing applications. These agents feature anthropomorphic avatars and customizable names, such as Dotty McDotface. The interface allows users to chat with the agent while following its actions in a separate window on a virtual machine. Dots are positioned more like workplace software that can interact with other applications, such as Blender and GIMP, granting access to the user's computer via the desktop ChatGPT app. Access to Dots is currently offered to users on OpenAI's highest-tier accounts, including the $100-per-month Pro subscription. Testing personal tasks revealed inconsistencies; for example, an attempt to schedule an installation appointment involved a step requiring manual mouse holding, and attempts to complete financial transactions were unsuccessful due to a lack of direct Stripe integration. Results varied across different tasks; some agents failed to find free options while others succeeded in scheduling appointments or performing file organization after being granted broader access to the user's computer.

The experience with the Dots agents highlights a tension between aspirational, highly personalized AI interfaces (like Meta Muse) and the practical constraints of enterprise-grade software integration. The performance gaps observed when interacting with external systems, such as handling security checks or third-party website protocols, suggest that current agent design is not universally robust across varied digital environments. The success in tasks requiring large contextual input—such as redesigning a personal website based on extended verbal instructions—demonstrates AI's potential for iterative, high-level creative execution when explicitly provided with extensive datasets. The pivot point occurs when the agent’s functionality shifts from being a simple personal shopper or informational tool to becoming a true operating layer within the user’s computational environment. This suggests that value is realized not in perfect personal service, but in the capacity for sophisticated workflow automation enabled by deep system access. The implicit trade-off—gated enterprise features versus functional reality—raises questions about whether imposing "workplace" structure inherently limits agents' utility when personal, low-friction tasks are introduced.