OpenAI is testing a new “Persistent mode” for Codex that could allow its AI coding agent to continue working on tasks until a user explicitly stops it. The feature has been spotted in Codex’s command-line codebase and is not yet broadly available. OpenAI has confirmed that it is testing the capability but said there are no immediate plans to launch it, making the feature a potential future addition rather than an announced product release.
The proposed mode would represent a significant shift in how Codex operates. Current agent sessions generally stop after completing a task or reaching time and usage limits, while Persistent mode appears designed for much longer-running work. Internal code reviewed by WIRED also points to a “proactivity” capability that could allow Codex to generate follow-up tasks for itself, continue working across sessions and potentially contact users when necessary. The development comes as OpenAI increasingly positions Codex as an agent that can handle complex work over extended periods.
OpenAI Tests Persistent Mode For Codex
The new feature was identified in OpenAI’s publicly visible Codex command-line code.
Persistent mode appears as an option related to the amount of reasoning, computing power and time Codex can use for a task. The code suggests that, when activated, Codex would continue working rather than automatically stopping after a relatively short period.
OpenAI has confirmed that the feature is being tested, but it has not announced a general release date.
Codex Persistent Mode At A Glance
| Particular | Details |
|---|---|
| Company | OpenAI |
| Product | Codex |
| Proposed feature | Persistent mode |
| Current status | Testing |
| Broad release | Not announced |
| Initial discovery | Codex command-line code |
| Operating model | Continue working until stopped |
| Potential capability | Proactive follow-up tasks |
| Cross-session work | Appears to be supported |
| User notifications | Potentially available |
| Main objective | Longer-running autonomous work |
The feature should therefore be treated as experimental. Its current presence in the codebase does not guarantee that every capability will eventually become available to users.
How Persistent Mode Could Change Codex
Codex is already designed to perform multi-step software-engineering tasks.
Users can give it a goal, allow it to inspect files, write code, execute commands, run tests and make changes. OpenAI has also introduced Goal mode, which lets users define an outcome and success criteria and allow Codex to keep working toward that objective.
Persistent mode appears to take this concept further.
Current Agent Workflow
User Gives Task
↓
Codex Analyzes Problem
↓
Writes / Modifies Code
↓
Runs Tests
↓
Reports Result
↓
Session Ends
Potential Persistent Workflow
User Gives Goal
↓
Codex Analyzes Problem
↓
Writes Code
↓
Runs Tests
↓
Finds New Issue
↓
Creates Follow-Up Task
↓
Continues Working
↓
Checks Results
↓
Repeats Until Stopped
The difference is that the agent would no longer treat the completion of the original request as the automatic end of its work.
‘Continue Working Until Put To Sleep’
The code associated with Persistent mode reportedly describes Codex as continuing to work until it is “put to sleep.”
That wording highlights the intended difference from conventional AI interactions.
Instead of waiting for another user prompt, the agent could remain active and continue making progress.
For software developers, this could be useful for projects where tasks naturally evolve as the agent discovers new problems.
Potential Long-Running Codex Tasks
| Task | Why Persistence Could Help |
|---|---|
| Large code refactoring | Multiple files may need changes |
| Bug investigation | New issues can emerge during testing |
| Test creation | Tests may reveal additional failures |
| Dependency upgrades | Compatibility issues can require follow-up work |
| Security audits | Findings may require several fixes |
| Documentation | Agent can continue through multiple sections |
| Performance optimization | Repeated testing may be required |
| CI/CD debugging | Failures can require several iterations |
Instead of requiring developers to repeatedly restart Codex, Persistent mode could allow the agent to manage these iterations itself.
Codex Could Become More Proactive
The most notable aspect of the proposed feature may be “proactivity.”
According to the code reviewed by WIRED, a Persistent mode agent could be instructed that its work is not necessarily finished after answering the user’s original request.
It could create additional tasks based on previous interactions and information it has about the user.
This would move Codex from a task executor toward a more autonomous digital worker.
Reactive Vs Proactive AI
| Reactive AI | Proactive AI |
|---|---|
| Waits for user instruction | Can identify follow-up work |
| Completes assigned task | Continues toward broader objective |
| Stops after response | Can remain active |
| User creates next task | Agent may create next task |
| Shorter sessions | Longer-running sessions |
| Prompt-driven | Goal-driven |
The distinction is important because an agent that can independently decide what to work on requires considerably stronger controls than a system that simply responds to individual requests.
Codex Could Work Across Sessions
The proposed proactivity functionality reportedly includes the ability to continue work across sessions.
That would allow Codex to maintain a longer-term understanding of a project rather than treating every interaction as an isolated request.
OpenAI has already been moving Codex in this direction.
The company’s own guidance on long-running Codex work emphasizes durable threads, preserved context, complex workflows and continuity across projects.
Long-Running Coding Workflow
Project Context
↓
Codex Thread
↓
User Goal
↓
Agent Work
↓
Results
↓
New Context
↓
Follow-Up Work
↓
Continued Progress
This could make Codex more useful for projects that take days or weeks rather than minutes or hours.
User Notifications Could Become Part Of Persistent Codex
The proposed system also appears to give the agent a tool for messaging the user without being explicitly prompted.
However, the internal instructions reportedly tell the agent to use this capability sparingly.
That limitation is important.
An always-active AI agent that sends frequent notifications could quickly become disruptive.
Instead, the intended model appears closer to an assistant that remains quiet while work is progressing and contacts the user when human input is actually useful.
Potential Notification Scenarios
| Situation | Possible Codex Action |
|---|---|
| Task completed | Notify user |
| Major test failure | Ask for guidance |
| Approval required | Request permission |
| Ambiguous requirement | Ask a question |
| New major issue discovered | Alert user |
| Work continues normally | Remain silent |
| Minor issue | Attempt resolution independently |
The balance between autonomy and interruption will likely be critical if Persistent mode is eventually released.
Persistent Mode Would Not Remove Permission Controls
The proposed feature does not appear to give Codex unlimited authority.
The internal instructions reportedly state that Persistent mode does not expand what the agent is permitted to do.
Changes outside the user’s own system would still require approval.
This is particularly important because a continuously running coding agent could otherwise make unintended changes or interact with external systems without the user’s knowledge.
Potential Safety Boundaries
Persistent Agent
↓
Existing Permissions
↓
User Approval Required For Restricted Actions
↓
Controlled Execution
The distinction is that persistence changes how long the agent works, rather than automatically expanding what it is allowed to do.
OpenAI Has Already Been Extending Codex’s Autonomy
Persistent mode would build on several recent Codex developments.
OpenAI has introduced Goal mode, computer use, remote access and mobile capabilities that allow users to monitor and steer Codex while work continues on connected machines.
The company has also emphasized longer-running workflows as an important direction for the product.
Codex Evolution
| Development | Significance |
|---|---|
| Basic coding assistance | Generates and edits code |
| Goal mode | Works toward defined outcomes |
| Computer use | Interacts with applications |
| Remote access | Allows work to continue while user is away |
| Mobile access | Lets users monitor active work |
| Long-running workflows | Preserves project continuity |
| Persistent mode | Potentially enables continuous operation |
This progression suggests that OpenAI is gradually moving Codex from a conventional coding assistant toward a more autonomous software-engineering agent.
Developers Could Leave Codex Working While Away
One of the biggest practical advantages of a persistent agent would be the ability to delegate work without remaining at the computer.
A developer could potentially assign a large task before leaving work and return later to review the results.
For example:
6:00 PM
Assign large refactoring project
↓
8:00 PM
Codex completes initial changes
↓
9:00 PM
Tests reveal additional issues
↓
10:00 PM
Codex investigates and fixes failures
↓
11:30 PM
Additional tests completed
↓
Next Morning
Developer reviews final changes
This is still a potential workflow rather than a currently available Persistent mode feature.
Persistent AI Agents Raise New Risks
The same autonomy that makes Persistent mode attractive also introduces additional risks.
A conventional chatbot generally stops after responding. A persistent agent can continue taking actions, which creates more opportunities for errors.
A mistake made repeatedly over several hours could have a much larger impact than a mistake made during a single interaction.
Risks Of Always-On AI Agents
| Risk | Potential Impact |
|---|---|
| Incorrect assumptions | Agent may pursue the wrong objective |
| Repeated errors | Mistakes can compound |
| Excessive resource use | Long sessions can consume significant compute |
| Unauthorized actions | Potential security consequences |
| Notification overload | User may receive unnecessary alerts |
| Context errors | Agent may misinterpret older information |
| Security vulnerabilities | Longer operation increases exposure |
| Goal drift | Agent could pursue unintended follow-up tasks |
This makes permission systems, monitoring and clear stop mechanisms particularly important.
Security Is Becoming More Important For Persistent Agents
The timing of the development is notable because OpenAI has recently highlighted the security challenges associated with increasingly autonomous AI systems.
Coding agents can already inspect repositories, execute commands and interact with development environments.
Giving an agent more time to operate increases the importance of sandboxing, permissions and human oversight.
For businesses, the question will not simply be whether an AI agent can complete a task, but whether it can do so safely over an extended period.
Security Layers For Persistent Coding
Agent
↓
Sandbox
↓
Filesystem Permissions
↓
Command Restrictions
↓
Network Controls
↓
Human Approval
↓
Audit Logs
A persistent Codex would likely need these layers to operate safely in professional environments.
OpenAI’s Codex Strategy Goes Beyond Coding
Although Codex is primarily associated with software development, OpenAI’s broader direction appears to be toward general-purpose agents capable of performing tasks across applications.
The company’s recent Codex updates have increasingly focused on computer use, remote access and long-running workflows.
This creates a potential path from:
Coding Agent
↓
Computer-Using Agent
↓
Long-Running Agent
↓
Proactive Agent
↓
General-Purpose Digital Worker
Persistent mode could therefore be important beyond programming if similar capabilities eventually appear in other OpenAI products.
Competition In Agentic AI Is Intensifying
OpenAI is not developing persistent agents in isolation.
Anthropic, Google and Meta are also investing heavily in AI systems designed to perform multi-step tasks with less human intervention.
The competition is shifting from who can generate the best single response to who can reliably complete entire workflows.
Agentic AI Competition
| Company | Broad Direction |
|---|---|
| OpenAI | Codex and increasingly autonomous agents |
| Anthropic | Claude-based agentic workflows |
| Gemini agents and computer-use capabilities | |
| Meta | AI agents across consumer and business products |
| Microsoft | Copilot and enterprise agents |
Persistent operation could become an important feature in this competition because users may increasingly judge AI systems by the amount of work they can complete independently.
The Bigger Picture
OpenAI’s experimental Persistent mode points toward a significant change in the role of coding agents. Instead of operating only when a developer provides a prompt, Codex could eventually remain active, identify follow-up work and continue pursuing a broader objective until the user stops it. The capability is still being tested and has not been broadly launched, so its final design and availability remain uncertain.
The development also highlights the central challenge facing autonomous AI: persistence increases both usefulness and risk. An agent that can work for hours without supervision could complete complicated software projects more efficiently, but it also needs strong permissions, monitoring and clear boundaries to prevent errors from accumulating. OpenAI’s existing investments in Goal mode, computer use, remote access and long-running workflows provide a foundation for this direction.
Looking Ahead
The biggest question is whether OpenAI ultimately turns Persistent mode into a broadly available Codex feature and how much autonomy it gives users. If released, the feature could allow developers to delegate larger engineering projects and let Codex continue testing, debugging and refining software without requiring a new prompt after every stage.
For OpenAI, Persistent mode could also become a test case for a much broader vision of always-on AI agents. The company will need to balance autonomy with compute costs, security, user control and reliability. If those problems can be solved, Codex could evolve from a coding assistant that responds to instructions into a continuously operating software-engineering partner
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
