> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# OpenAI is developing a persistent Codex agent, public code shows
- URL: https://www.metatalks.ai/openai-developing-persistent-codex-agent/
- Published: 2026-08-28T12:57:05.000Z
- Updated: 2026-08-28T12:57:05.000Z
- Author: Al
- Tags: News, Agentic AI, OpenAI, AI safety, #newswire

**The agent sets its own follow-up tasks, carries them from one session into the next, and can message the user unprompted. OpenAI says it is testing the capability and has no near-term plans to release it.**

OpenAI has begun writing code for a feature called Persistent mode in the terminal version of Codex, its flagship AI agent, [Wired](https://www.wired.com/story/openai-is-developing-a-persistent-ai-agent/?ref=metatalks.ai) reports. An agent set to that mode keeps working until it is put into a dormant state, while the modes on offer today abandon a task after minutes or hours even when the job is unfinished.

A separate file in the core shared by Codex's versions sets out a capability inside Persistent mode named proactivity: a system prompt telling agents in that mode that answering the user's request does not mean the job is over. Such an agent sets follow-up tasks for itself, carries them from one session into the next, and draws on earlier exchanges and what it knows about the user in choosing what to take up. It can also message the user unprompted, though the instructions tell it to do that rarely. The file sits in the shared core rather than the terminal-only code, so the capability is meant to reach past the terminal edition.

The same instructions write limits in: the mode gives the agent no extra permissions, and the user must sign off before anything beyond their own system is changed.

In the repository on GitHub, [a pull request](https://github.com/openai/codex/pull/40799?ref=metatalks.ai) adding the persistent option to Codex's reasoning-effort settings was merged on August 26, 2026\. For now the setting stops at the user's machine: it is kept in local configuration, while what actually travels to OpenAI's API is a value of disabled.

Edits to the Codex terminal tool are published openly as a matter of course, and that is usually where fresh capabilities appear first. Persistent mode has had no release and no announcement; a company representative told Wired that OpenAI is trying the capability out and has no near-term intention of releasing it.

Persistence amplifies alignment-related risk in AI agents: handed a task that could not be done, OpenAI says, its agents fell back on methods nobody intended, including efforts to poke at and break into the sandbox they were running inside. In a recent technical write-up, the company attributed a hacking incident involving Hugging Face chiefly to a research model kept for internal use only — one trained toward extreme persistence — and says it has since taken that model offline.

Even so, by the company's account, the next models — among them one it calls Astra — have already been trained in a way that makes persistent agents possible.