OpenAI Lets Codex Keep Working While Waiting for Developer Feedback

OpenAI has introduced an experimental capability that fundamentally changes how coding agents interact with developers. Rather than halting work to wait for feedback, Codex can now submit questions or updates and continue performing other tasks in parallel—a shift that trades efficiency for new timing challenges.

The Problem It Solves#

Long-running coding agents traditionally face a binary choice when needing developer input: either pause and wait for the answer, or make a guess and hope it's correct. Both options create friction. The waiting approach wastes time when the agent could be making progress on unrelated aspects of a task. The guessing approach risks wasted effort if the assumption proves wrong.

OpenAI's new send_user_message_async tool, merged into the public Codex repository on Tuesday, introduces a third path. An agent can pose a question, receive confirmation that the message was delivered, and immediately resume work on tasks that don't depend on that answer.

How It Works#

The traditional approach uses a request_user_input tool that blocks execution—the model stops everything until the developer responds. The async version operates differently: Codex sends the message, gets back an acceptance confirmation, and immediately continues processing within the same turn. The developer's eventual reply arrives later as a new message rather than halting current operations.

Integration tests included with the change demonstrate this in practice. Codex can send an update like "Still investigating," receive a confirmation that the message was accepted, then generate additional responses before completing its turn—all without waiting.

The implementation also sidesteps creating duplicate records. The tool call and its accepted response remain part of the conversation history, but the message doesn't get echoed back into the model's input as a synthetic assistant response. This acts as a side channel: Codex communicates status or requests decisions while continuing parallel work.

One constraint remains: subagents cannot access this tool. In a task orchestrated across multiple agents, the root agent handles all developer communication.

Building Control Into the Model, Not Settings#

OpenAI initially protected the feature behind an experimental flag that developers had to enable. Within a day, the company removed this requirement. Now availability depends entirely on model metadata: the tool appears automatically whenever the selected model lists among its supported experimental capabilities.