No More "Wait, Stop, Start Over": Codex 0.163 Makes Steering a Running Agent Mid-Response the Default

OpenAI's coding agent "Codex CLI" released version 0.163.0 on October 16. The headline feature is real-time steering, which lets you send additional instructions to an agent while it is still responding and change its course on the spot. What used to be an optional setting is now enabled by default. Alongside it come changes that affect the day-to-day feel of the tool, such as parallel execution of read-only tools and a searchable settings screen.

You Can Rephrase Even After the Reply Has Started

Interacting with an agent has traditionally been a sequential process: give an instruction, wait for the reply, and if something bothers you, stop it and start over. Typing an additional message partway through usually meant waiting until the current response finished (it got queued), or interrupting to stop it first and then resubmitting.

In 0.163.0, a follow-up you send while Codex is responding takes effect immediately, letting you correct course without halting the generation in progress. The official release notes explain that "sending a follow-up while Codex is responding lets you steer it instantly," and they explicitly state that on an interrupt, any running Code Mode cell (including the nested tool calls running inside it) is also stopped. The behavior can also be enabled from settings on the ChatGPT desktop side; what was previously opt-in is now the default behavior.

Put simply, the move of "cutting in from the side on a running agent to change its course" has become standard.

Four Changes That Aren't So Small for the Everyday Feel

Beyond steering, a bundle of updates reshapes the feel of the parts you reach for repeatedly.

• Parallel execution of read-only tools: Read-only operations that don't interfere with one another—such as referencing skills, memory, and conversation history—can now run simultaneously. Preliminary lookups that used to wait in line one by one get handled all at once.
• A searchable settings screen: /config is now a panel that lets you search settings across categories, with supported items applied on the spot. It also shows whether an item requires a restart.
• A changeable leader key: /keymap lets you set the leader key that serves as the starting point for shortcuts (the default is Ctrl+X).
• Restarting the resident server: From /daemon you can restart the managed background server, which warns you about work that may be interrupted and then reconnects automatically.

On top of this, finer gaps have been patched—for example, permissions and network approvals are now restored to their original state when you resume a saved conversation.

It Was a Problem of "Timing," Not "Wait Time"

What real-time steering changes is not the speed of processing itself, but the timing of the back-and-forth between human and agent. Until now, even when you noticed the output had started heading off in the wrong direction, your only options were to wait for it to finish or to interrupt, stop it, and rework the instruction. Waiting piles up wasted tokens and time; stopping breaks your train of thought.

Being able to rephrase midway lets you step off this either-or. In the middle of a long implementation or investigation, you can now slip in "don't touch that file" or "show me the plan first before anything else," and the interaction shifts from sequential round-trips toward a continuous flow closer to conversation. Among developers who have long recommended steering, the shift to real time has prompted remarks that "there's no longer any reason to stockpile messages while a response is running." This is a matter of efficiency, but it is also a shift toward using agents not by "handing everything off" but by "running alongside them and nudging the rudder in small increments."

Where It Pays Off in Team Operations

From a work standpoint, the fact that the cost of course correction drops is no small thing. The rework from a derailed generation (waiting and redoing) translates directly into token billing and idle time for agent usage. Being able to correct midway reduces how often you let it run to the end only to throw the whole thing away. Parallelizing read-only tools also helps cut wait time on investigation tasks that involve many lookup round-trips.

Making settings searchable and tuning shortcuts may seem minor, but the more a team uses agents daily with multiple people, the more the learning cost of "where to configure what" accumulates. How easy settings are to find ties directly to how well the tool takes hold during rollout.

The Caveats Attached to the Default Switch and Daybreak

At the same time, it's worth keeping in mind the cautions that lie behind the convenience.

First, the fact that steering is now enabled by default is itself a change. If you have automation or scripts built on the assumption that messages get queued during a response, the behavior may change, and some point out that a review is warranted. On the parallel execution of read-only tools as well, some commentary cautions that the interplay with permission approval has not been spelled out in detail. When incorporating this into unattended operation, it's safest to do a one-time check on the premise that the default behavior has changed.

In addition, there is an item called Daybreak that is now described as "enabled by default for eligible accounts" (previously it was said to be handled by explicitly enabling a flag like --enable cli_daybreak). The release notes say "enabled by default for eligible accounts; with codex exec you can also select it if you have an OpenAI API key, to the extent the server allows," but no explanation has been published for what Daybreak actually refers to. It's also unclear what the conditions for being "eligible" are, or what happens if the server does not provide access. Since the substance is unpublished, for now the reasonable stance is to note only the fact that "there is an unexplained feature that has been enabled by default" and to avoid excessive speculation.

References: Release 0.163.0 · openai/codex (GitHub) / Codex Updates by OpenAI (Releasebot) / Codex CLI 0.163.0 — steer mid-response (ai-tldr.dev) / Codex CLI 0.163.0: Daybreak by Default, Instant Interrupts (gptmap.org)