Software · 28 August 2026 · 4 min read
Codex Goes Tireless: OpenAI Tests Persistent, Proactive Mode for AI Agents
In brief: OpenAI is developing a new operating feature for Codex called «Persistent mode», enabling the AI coding agent to run indefinitely until explicitly put to sleep by the user. The architecture includes proactive system directives that let the model self-assign follow-up tasks and ping users across sessions. The initiative marks a decisive shift from reactive coding assistants to autonomous, long-running digital collaborators.
by Team Mocchi's
The paradigm of software development assistants is poised for a structural shift. Until now, interactions with autonomous coding agents have largely followed a reactive loop: a developer issues a prompt, the model executes a sequence of operations for a few minutes, and once its execution budget or context limits are reached, it hands control back to the human operator. OpenAI is now looking to break through this runtime barrier with a system designed to operate continuously in the background.
As revealed in a codebase analysis by WIRED, the company led by Sam Altman has added code for an experimental feature called «Persistent mode» to its Codex command-line interface. Currently under active development and testing, the feature allows the agent to keep working across sessions until the user explicitly puts it to sleep.
Beyond Timeouts: How Persistent Mode Operates
Within the Codex repository, Persistent mode appears inside the «reasoning effort» selector—the configuration parameter controlling the compute budget, token allocation, and processing time an AI model may dedicate to an issue before taking action. In standard modes, agents cease execution after a set timeout, even if a larger task remains incomplete.
When Persistent mode is engaged, OpenAI's internal system directives state that Codex will «continue working until put to sleep.» This represents one of the most computationally demanding settings built into an agent product, engineered to tackle large-scale refactorings, multi-module migrations, or comprehensive test suite generation that span hours or days of continuous iteration.
Proactivity and Autonomous Backlog Management
The fundamental innovation lies in the «proactivity» engine tied to Persistent mode. According to configuration files examined in the repository, the agent is governed by a dedicated system prompt: its task is not marked as finished once the initial prompt is resolved. Instead, the AI is instructed to proactively formulate follow-up tasks for itself.
By leveraging past interactions, workspace context, and accumulated user preferences, the agent determines which parts of the codebase require optimization, bug fixes, or missing documentation. The architecture also allows the model to initiate unprompted messages to the developer when reaching key milestones or requiring high-level clarification, accompanied by strict instructions to minimize alert fatigue.
Safety Guardrails and Execution Boundaries
Because the proactivity logic sits inside the core framework shared across Codex rather than exclusively in terminal-specific files, the capability is expected to extend beyond command-line tools into desktop interfaces and enterprise products such as ChatGPT Work. Given the risks inherent to long-running autonomous processes, OpenAI has embedded explicit safety boundaries into the system.
The system instructions explicitly clarify that enabling Persistent mode does not broaden the agent's privilege envelope. Any action involving modifications outside the user's defined local environment or touching protected network resources still mandates explicit human confirmation.
Mocchi's take
The shift toward persistent, proactive AI agents fundamentally alters how engineering teams approach software lifecycle management. For tech businesses and software organizations, this transition opens up immense efficiency gains in technical debt remediation, continuous regression testing, and repository hygiene by offloading them to asynchronous, tireless background workers. However, continuous autonomy demands unprecedented architectural rigor: to prevent runaway token costs and unexpected drift, organizations must implement robust isolation sandboxes, deterministic CI/CD validation gates, and strict branch protection policies before letting autonomous agents iterate unchecked on production codebases.