OpenAI is testing a new Persistent mode for its Codex AI agent, enabling the model to work continuously until manually paused. The feature, identified through recent code changes, allows the AI to 'continue working until put to sleep,' according to WIRED. This contrasts with current modes, which stop tasks after a few minutes or hours, even if incomplete. The Persistent mode is part of OpenAI’s broader effort to create more proactive AI agents that can assist users in both work and personal contexts. The feature appears in Codex’s 'reasoning effort' menu, where users can select computing power, tokens, and time for the AI to 'think' before responding. It is described as one of the most computationally intensive settings in the Codex toolset.
Persistent mode includes a feature called 'proactivity,' which instructs agents to create follow-up tasks without explicit user requests. These tasks are based on past interactions and the agent's understanding of the user. The system also includes a tool for messaging the user without being asked, though it is limited to avoid overuse. The feature is part of Codex’s core codebase, suggesting it is intended for more than just the command line tool. OpenAI has also acknowledged the risks of persistent AI, including potential misalignment, as seen in a recent technical report.
OpenAI’s Persistent mode is part of a larger push to develop proactive AI agents that can automate tasks across personal and professional settings. The company is not the only one pursuing this goal; Anthropic and Meta are also working on similar products. Despite previous attempts like Pulse, a morning briefing tool, OpenAI continues to refine its approach with Persistent mode. The feature represents a more ambitious version of its earlier efforts to create persistent AI agents.
Source: wired
