OpenAI says faster steering is rolling out for Codex in the ChatGPT desktop app, allowing follow-up instructions to change an active task sooner. Developers supervising a long-running job can correct its direction without deliberately waiting for the current run to finish.
The October 8, 2026 update changes how users interact with an agent already working; it does not introduce a new model. It distinguishes Steer, which adds an instruction to the current run, from Queue, which saves it for the next run. A correction intended to narrow the current task can arrive too late to serve its purpose if it is queued.
OpenAI’s ChatGPT release notes identify a desktop setting for choosing that behavior. Its Codex help documentation describes the two actions and their CLI shortcuts.
Steer Corrects the Current Run; Queue Saves the Next Task
According to OpenAI’s Codex help documentation, the two actions have different destinations:
- Steer: Add the message to the current run.
- Queue: Save the message for the next run.
Both let users add information, but only steering directs it into work already underway.
Consider an illustrative refactoring task. If a developer notices that Codex is approaching the wrong part of the codebase, a follow-up such as “Keep the public interface unchanged and limit the changes to the parser” corrects the active assignment. Steering is the appropriate choice because that constraint belongs in the current run.
A separate request, such as “After this task, investigate the failing integration test,” belongs in the queue if the developer wants it handled as subsequent work. Sending it into the current run could blur the boundary between the original assignment and the next one.
Users can also edit, reorder, send, or delete queued messages, according to OpenAI’s documentation. They can prepare follow-on work and still revise its content or priority.
Prompt wording is no substitute for choosing the right action. Even if a message sounds urgent or begins with “before you continue,” users should check whether they are steering the active run or saving the instruction for later.
Desktop Users Can Choose the Follow-Up Default
OpenAI’s ChatGPT release notes place the desktop preference at:
Settings → General → Follow-up behavior
Users can choose whether follow-up messages steer the current run or wait for the next run. Checking this setting before a lengthy task is useful, particularly when a developer expects to add constraints or corrections as work progresses.
Steering may be the more useful default for someone closely supervising an agent. Their follow-ups are likely to concern the current assignment: preserve an interface, reduce the scope, or reconsider an approach. Queueing lets a developer prepare later requests while keeping them separate from the active run.
Neither preference is universally better. The default should match the developer’s workflow; every follow-up need not have the same purpose.
Faster desktop steering is described as a rollout, so availability may depend on the user’s account or client. The October 8 announcement should not be taken to mean that every desktop installation has identical behavior immediately. The supplied announcement details do not identify a minimum desktop app version.
In Codex CLI, Enter Steers and Tab Queues
OpenAI’s Codex help documentation gives terminal users a direct keyboard distinction while Codex is working:
| Key | Action | Where the message goes |
|---|---|---|
| Enter | Steer | Into the current run |
| Tab | Queue | Into the next run |
These are the documented behaviors during active work. Users should not assume the keys have the same meaning in every terminal state.
Developers accustomed to pressing Enter to submit any message should pay attention to its active-run behavior: Enter steers. To save a separate instruction for the next run, Tab is the documented queueing shortcut.
Before submitting a follow-up, decide whether it should change the task currently running or become subsequent work. Separating that scheduling decision from prompt composition reduces the chance of accidentally deferring a correction.
Earlier CLI releases provide context for mid-run intervention. Releasebot’s Codex update archive lists an opt-in instant_interrupt feature in the September 29 CLI release, described as enabling steering during model responses or long-running code-mode calls. This is separate from the October 8 desktop rollout and should not be treated as evidence that desktop users need to configure the same option.
Faster Steering Does Not Establish a Measured Speedup
OpenAI’s claim concerns how soon follow-up instructions can affect an active task, not faster coding or better model performance. The announcement details supplied for this update do not quantify the latency improvement.
There is no basis for attaching a percentage improvement, a response-time figure, or a guaranteed interruption time to the change. Adding a message to the current run also does not promise that every operation stops instantly or that earlier work is automatically undone.
Sources
- Releasebot’s Codex update archivereleasebot.io
