The race for autonomous desktop agents is accelerating rapidly. Following the launch of Meta's personal AI agent Muse, which quickly climbed to the top of major app store charts, OpenAI has responded by overhauling its agent infrastructure. At the center of this product rollout are Dots, persistent background agents hosted on isolated cloud virtual machines. Equipped with a dedicated browser and read permissions for connected enterprise workspaces such as Slack, Microsoft Teams, and ChatGPT, these instances operate continuously. Driven by GPT-6 Astra, Dots handle complex tasks including invoice audits, data analytics, and routine software bug fixes proactively without requiring continuous user prompts.
In parallel, OpenAI introduced Computer History directly within the ChatGPT interface. This capability analyzes desktop activity locally on client machines. By tracking active user interactions, the system reconstructs interrupted tasks and identifies opportunities to automate recurring processes. The feature marks a strategic shift away from static prompt-response interfaces toward continuous, contextual workflow assistance.
For external developers, OpenAI deployed the new Decisions API. The interface was engineered within a single week to counter specialized tools such as Jev. It consolidates several critical capabilities into a unified endpoint: asynchronous tool calling, mid-turn steering for dynamic redirection during model inference, connection pre-warming, state compaction, and native WebSocket connectivity for low-latency feedback loops.
Appearing on the Latent Space podcast, Ari Weinstein, head of the Computer-Using-Agent team at OpenAI, alongside Nikunj Handa, detailed the underlying architectural transition. Weinstein argued that practical agentic systems have moved well beyond the purely vision-driven paradigm. Rather than relying solely on analyzing pixel arrays and emulating cursor clicks across a display, modern agents adopt a layered hybrid approach.
Modern implementations combine captured screenshots with operating system accessibility trees, the browser Document Object Model, the Playwright automation framework, and ad-hoc code execution. This combination yields dramatic reductions in execution latency and significantly curtails interaction errors compared to earlier computer-use implementations. To provide reproducible evaluations for this architecture, the OSWorld 2.0 and DeepSWE 1.1 benchmarks were released to the research community.
OpenAI's latest releases reflect intense commercial pressure following Meta's Connect announcements and the growing popularity of competing workflow automation agents. As leading tech providers vie for control over the primary desktop workspace, always-on agents signal a decisive migration away from reactive chatbots toward persistent digital colleagues.