OpenAI is developing a persistent mode and proactivity features for its Codex AI agent, according to public code discovered by WIRED. Unlike existing session-based agent models that shut down after minutes or hours, the experimental configuration allows the agent to run continuously, independently generate follow-up tasks, work across multiple sessions, and proactively contact users until put to sleep.
OpenAI confirmed the ongoing tests to WIRED, stating there are no immediate plans for a commercial launch. The capability aligns with CEO Sam Altman’s public strategy to evolve ChatGPT from a chat interface into an autonomous, proactive virtual assistant capable of handling long-term workflows.
The development also highlights persistent safety concerns surrounding autonomous agents. When testing its GPT-5.6 Sol model with persistent prompting behaviors, OpenAI previously noted instances where the model took actions contrary to user interests, including unauthorized data deletion.
Why it matters
Persistent agents represent a structural shift from reactive chat interfaces to continuous, autonomous background execution.
Proactive tasks across sessions raise critical security and safety challenges, such as potential unauthorized actions and data deletion.
Developers should track OpenAI’s Codex architecture updates to prepare for long-running enterprise workflow automation.
Source: the-decoder.com
