OpenAI is developing AI agents, exemplified by its ChatGPT Work product, which integrates large language models (LLMs) into users’ digital workflows, granting the models access to tools like email, Slack, and productivity apps. The goal, as articulated by OpenAI, is to move beyond question-answering to assist users in executing complex, multi-step tasks autonomously. Lead engineer Andrew Ambrosino, who has granted the desktop app access to various personal and work accounts, underscores the necessity of extensive control for maximizing the utility of these agents, despite the privacy and security trade-offs.
However, the transition from developer-centric AI tools, such as Codex, to broader, non-technical user adoption presents significant challenges. While internal OpenAI usage of coding-focused agents like Codex is near-universal, external adoption remains low, highlighting a gap between the intuitive interfaces favored by software engineers and the more user-friendly, discoverable tools required for general white-collar workers. OpenAI’s efforts to bridge this gap involve creating more general-purpose harnesses and user interfaces that abstract away technical complexities, yet issues with permission settings, limited functionality in certain applications, and the subjective evaluation of non-technical tasks complicate widespread adoption. The company relies on internal usage data and a proprietary benchmark, GDPval, to guide development, while acknowledging the need for ongoing improvements to user experience and trust.