OpenAI Is Creating a Long-lasting AI Assistant

OpenAI is working on an advanced, highly engaged version of its leading AI agent, Codex, according to information obtained by WIRED.

Recently, OpenAI began incorporating code for a new “Persistent mode” feature into its command line iteration of Codex, based on changes to the product’s code base reviewed by WIRED. Modifications to the Codex command line tool are publicly accessible by default, with new functionalities often appearing there before transitioning to OpenAI’s other offerings, like the Codex desktop application and ChatGPT Work.

Although Persistent mode has not been widely implemented or announced, it may be forthcoming. An OpenAI representative confirmed to WIRED that the company is testing this feature, but indicated that there are no immediate plans for its release.

“OpenAI fosters a very grassroots culture, exploring various avenues in the open source repository which serves as a sort of our collective playground,” stated Thibault Sottiaux, OpenAI’s head of core products, in a statement to WIRED.

This feature seems to be OpenAI’s most recent effort to develop AI agents that people find genuinely useful. OpenAI, along with Anthropic and Meta, are in a race to deliver general-purpose agent products that assist individuals in automating tasks in both professional and personal realms, like submitting expense claims or scheduling medical appointments. Currently, the primary users of AI agents are largely software developers, but Silicon Valley sees potential for the technology to become a significant revenue stream with a much wider customer audience.

Persistent mode is featured in Codex’s “reasoning effort” menu, where users can choose the desired level of computational power, tokens, and time to be allocated for an AI model to “think” before responding to a prompt. It appears to be among OpenAI’s most resource-demanding settings. When Persistent mode is activated, the codex is programmed to “continue operating until instructed to stop.” This significantly contrasts with current modes, which terminate their tasks after a set period, even if they remain unfinished.

Additionally, within the code base, OpenAI elaborates on a function in Persistent mode termed “proactivity.” This seems to be a system prompt designed for agents in Persistent mode, indicating that their responsibilities aren’t complete once they fulfill a user’s request. Instead, the agent is directed to proactively generate follow-up tasks. This agent can address those tasks over multiple sessions, leveraging past user interactions and “knowledge of the user” to determine its next focus. It also possesses a capability to communicate with the user without solicitation, though it is instructed to do so sparingly.

The guidelines also define boundaries for the agent, as indicated in the file. The agent is informed that Persistent mode does not expand its capabilities and that any alterations outside the user’s system necessitate prior user approval—seemingly aimed at mitigating the potential risks of a persistent AI agent. The file is located in the core of Codex rather than the code specific to the command line, suggesting that the proactivity feature is meant for broader applications than just the terminal tool.

In recent podcasts, interviews, and private meetings with investors, OpenAI CEO Sam Altman has expressed his ambition to transform ChatGPT into a proactive, always-active AI agent. OpenAI aims for these advancements to enhance the usage of its most sophisticated AI models, which are currently employed by only a small portion of ChatGPT’s overall user base.