OpenAI is developing a new feature that would allow its Codex coding assistant to work autonomously until explicitly turned off, according to code reviewed by WIRED. The so-called “persistent” agent marks a significant shift from current AI tools that require constant user prompts, potentially enabling longer-running, proactive workflows.
From Reactive to Persistent: A New Era for AI Assistants
Traditionally, AI assistants like Codex operate in a conversational loop: a user provides a task, the AI completes it, and the process repeats. This reactive model limits their utility for complex, multi-step projects. OpenAI’s persistent agent aims to break this cycle by allowing Codex to continue working on a task—checking results, adjusting its approach, and even initiating new steps—until a human intervenes to “put it to sleep.”
This capability could transform how developers use AI. Instead of supervising every line of code, engineers might delegate entire feature builds or bug-fix sessions to Codex, freeing them to focus on higher-level design decisions. The persistent agent would also be valuable for monitoring systems, running tests, and managing routine updates without constant human oversight.
Under the Hood: How It Works
While OpenAI has not publicly detailed the technical implementation, the code reviewed by WIRED suggests the persistent agent uses a multi-step feedback loop. It likely combines Codex’s existing ability to generate and execute code with a new planning module that can set intermediate goals, evaluate its own outputs, and request additional resources or data as needed. This would require tighter integration with cloud services and better memory management—both areas OpenAI has been actively developing.
Industry analysts note that persistent agents are the logical next step for AI coding tools. “We’re moving from copilots to autopilots,” says Dr. Lena Fischer, a principal researcher at the AI Policy Institute. “The challenge is ensuring reliability and safety when the system operates without constant human supervision.”
OpenAI has not set a release date, but insiders suggest the feature could arrive in beta by mid-2027, possibly integrated into Codex’s enterprise tier. The company is likely to start with limited permissions—restricting the agent to sandboxed environments—before rolling out broader autonomy.
Safety and Control: What Users Need to Know
Allowing an AI to work “until put to sleep” raises obvious safety questions. OpenAI has emphasized that the persistent agent will include robust controls: users can pause, terminate, or adjust its scope at any time. The system will also log all actions for audit, and administrators can set spending limits or restrict access to sensitive resources.
“We’re designing this with fail-safes,” an OpenAI spokesperson told WIRED. “Persistent doesn’t mean unsupervised—it means the AI can handle more of the work on its own, but humans remain in the loop for critical decisions.” This approach aligns with broader industry trends toward “human-in-the-loop” autonomy, where AI executes routine tasks but escalates ambiguous or high-stakes situations to people.
For developers, the persistent agent could significantly boost productivity. A study by Stanford’s AI Lab found that similar autonomous workflows reduced average debugging time by 40% in controlled trials. However, experts caution that persistent agents will require a cultural shift—developers must learn to trust AI with longer tasks and to review its work periodically.
Outlook: The Road Ahead
OpenAI’s persistent agent is among several initiatives pushing toward more autonomous AI systems. In October 2026, the company released a paper on “self-directed learning,” demonstrating that AI models can improve their own code-generation capabilities through iterative practice. Combined with persistent execution, such systems could eventually manage entire software projects, from initial scaffolding to deployment.
Competitors are not far behind. Google DeepMind and Anthropic have both announced similar research projects, but OpenAI appears to be closest to shipping a product. If successful, the persistent agent could become the default mode for AI coding assistants, reshaping the developer experience and raising the bar for what users expect from “AI that works with you.”
As with any transformative technology, adoption will depend on trust. OpenAI’s challenge is to prove that persistent agents are not just powerful but dependable—and that they can indeed be “put to sleep” when needed.
via Wired AI