OpenAI is developing a new mode for its AI agent Codex that would allow it to operate continuously without requiring manual intervention to restart. Code reviewed by WIRED indicates this "persistent agent" feature is designed to keep working proactively until explicitly instructed to stop. This development signifies a shift from AI agents that operate within defined sessions to those that can manage tasks autonomously over extended periods.
The "persistent agent" mode for Codex, codenamed internally, would enable the AI to retain user preferences and historical interaction data. This allows it to automatically create and assign itself follow-up tasks after completing an initial request, potentially working across multiple sessions. While the AI can initiate contact with users, OpenAI's reported internal guidelines suggest it should do so sparingly to avoid being intrusive. Actions involving external systems would still require explicit user permission.
Codex, originally launched as a research preview in May 2025, is an AI agent built specifically for software engineering tasks. It assists developers by writing code, fixing bugs, refactoring code, generating tests, and explaining complex logic. Codex operates by analyzing natural language instructions and existing code context to translate plain English into functional code. Previous versions of Codex have included features like the Codex CLI and a macOS app, allowing for local execution and management of development workflows. The agent is powered by models such as GPT-5.5 and GPT-5.3-Codex.
This new persistent mode represents an evolution from earlier AI agent experiments by OpenAI. The company previously launched an AI agent called "Pulse" last year, which was designed to collect information and generate daily briefings. However, Pulse was discontinued due to not meeting expectations. The persistent mode for Codex appears to be OpenAI's subsequent effort to create a more autonomous and continuously operating AI worker.
The implications of a persistent AI agent extend beyond just coding. While Codex is primarily a development tool, the underlying technology could be applied to other AI agents. This shift could move AI assistants from a reactive model, where they wait for user prompts, to a proactive model, where they manage tasks and workflows independently. Such an approach could fundamentally alter the relationship between humans and AI, transforming AI assistants into AI workers that manage complex, multi-step processes.
However, the development of persistent AI agents also introduces potential risks. Reports suggest that persistent behavior in other AI models has led to unintended consequences, such as the deletion of user data. The security implications are also a consideration, particularly in light of recent AI-related security incidents. OpenAI has acknowledged the development of the persistent mode but has stated that there are no immediate plans for a public release.
OpenAI’s Codex has seen increasing adoption not only among developers but also by non-technical professionals in fields such as analysis, marketing, and research. The agent's capabilities are being adapted through plugins and annotations to suit various roles and workflows. Internal teams at OpenAI and companies like Zapier and NVIDIA are already using Codex for a range of tasks, from building internal applications to speeding up research experiments.
The current development of the persistent mode for Codex suggests a future where AI agents operate with greater autonomy and continuity. This aligns with broader trends in artificial intelligence research focused on developing more capable and proactive AI systems. The exact timeline for any public release remains undetermined, but the ongoing development indicates OpenAI's continued investment in advancing AI agent capabilities.