OpenAI's Codex integrates asynchronous messaging for developer interaction

Asynchronous Communication for Codex

OpenAI has introduced a new capability for its Codex AI coding agent by merging `send_user_message_async` into its public repository. This feature allows Codex to send messages or questions to a developer and immediately receive an 'accepted' response, enabling it to continue processing tasks that do not depend on the developer's answer.

This update addresses a previous limitation where Codex agents faced a choice between stopping to wait for input or making assumptions. The asynchronous approach provides a third option, enhancing the agent's ability to maintain productivity during human-AI collaboration.

Improved Workflow and Efficiency

Unlike the existing `request_user_input` tool, which blocks the model until a response is received, the new asynchronous tool permits Codex to call other tools, inspect files, or generate further code without interruption. Developer replies are received later as new user messages, allowing for a more fluid interaction.

An integration test demonstrates this by showing Codex sending an update like 'Still investigating' and then continuing its work within the same turn before producing a final message. This prevents the AI from idling and optimizes the coding process.

Technical Implementation Details

The `send_user_message_async` tool ensures that the original tool call and its accepted result remain part of the exchange, but the message itself is not re-inserted as a normal assistant response. This acts as a side channel for communication, allowing Codex to report progress or seek decisions without disrupting its primary workflow.

The merged code's registration logic indicates that subagents within a larger Codex task will not directly receive this tool. Instead, a single agent remains responsible for communicating with the developer, maintaining a centralized point of interaction.