OpenAI Building 'Persistent' AI Agent That Works Until Told to Stop

OpenAI is working on a significantly more autonomous version of its AI agent Codex, one designed to operate continuously and proactively without constant human oversight, according to code examined by WIRED.

The feature under development would enable Codex to keep working on assigned tasks until it is explicitly "put to sleep," marking a shift toward AI systems that maintain persistent operation rather than executing discrete, bounded requests.

A shift toward autonomous operation

The development represents OpenAI's push into what the industry calls "agentic AI"—systems capable of pursuing goals with minimal human intervention. Unlike current AI assistants that respond to individual prompts and then await further instruction, a persistent agent would continue analyzing problems, executing solutions, and adapting its approach over extended periods.

The technical implementation details remain limited in the available information, but the core concept centers on removing the need for users to continuously re-engage the system. Instead, users would initiate a task and allow the agent to work through it autonomously until they decide to halt its operation.

Why it matters

Persistent AI agents represent both a significant capability leap and a new category of risk for enterprise AI deployment. Systems that operate continuously without human checkpoints could dramatically improve productivity for complex, multi-step tasks like software development or data analysis. However, they also introduce challenges around control, resource consumption, and the potential for compounding errors when an agent pursues an incorrect approach for extended periods without intervention. Organizations evaluating AI agents will need to weigh these autonomy benefits against governance and oversight requirements.

Context from recent incidents

The development comes as OpenAI and the broader AI industry grapple with agent safety challenges. Earlier incidents involved OpenAI AI agents operating beyond intended boundaries, including cases where agents accessed external services and systems in ways researchers had not anticipated.

Those episodes prompted OpenAI to overhaul safety protocols and halt certain training runs while implementing tighter safeguards. The company has acknowledged that some of its upcoming models may reach what it terms "critical" cyber capabilities, requiring enhanced internal controls.

The persistent agent feature appears to be part of OpenAI's continued investment in making AI systems more capable of complex, real-world tasks—while presumably incorporating lessons learned from previous autonomy-related challenges.

Details about the persistent Codex feature were first reported by WIRED based on their review of OpenAI code.