October 8, 2026 — OpenAI is rolling out an update to how Codex handles follow-up messages during an active task in the ChatGPT desktop app. The update is designed to make instructions sent while Codex is already working take effect more quickly, helping users adjust a task without necessarily restarting the entire workflow.
According to the reported release-note update, the desktop app includes a control under General → Follow-up behavior for choosing whether a new message should steer the current execution or be queued for the next step. The change is being rolled out, so availability may vary.
What is Codex task steering?
Task steering means sending additional guidance while an agent is already carrying out work. Instead of waiting for a long operation to finish, a user can clarify a requirement, change priorities or correct the direction of the current task. The agent's ability to incorporate the instruction depends on its current execution state and the selected follow-up behavior.
Steer now versus queue for later
The two behaviors solve different problems. Steering the current task is useful when an instruction should influence work already underway—for example, correcting the scope of an audit. Queueing a follow-up is useful when the existing operation should complete before the agent handles another request. Neither option should be interpreted as a guarantee that every already-executed action can be undone.
Why faster follow-ups matter
Long-running coding and research tasks often involve intermediate discoveries. A user may realize that a target folder is wrong, a test should be added or a particular source should be excluded. Faster steering reduces the friction of course corrections and can make supervised agent workflows more practical.
Examples for SEO teams
- Technical audits: ask Codex to exclude staging URLs or expand a check to another sitemap while it is processing.
- Crawling workflows: adjust analysis scripts or filtering logic when unexpected URL patterns appear.
- Structured data validation: add a schema type or a specific validation case before the agent finishes its work.
- Reporting: change output fields or request additional segmentation during an ongoing analysis.
These are potential workflow applications, not a claim that Codex automatically operates a production crawler, accesses a website or modifies external systems without appropriate tools and authorization.
How to choose the behavior
In the ChatGPT desktop app, check Settings → General → Follow-up behavior when the feature is available. Choose the behavior that matches your workflow: intervene in the active task when the change is urgent, or queue the message when preserving the current run is more important. The precise wording and location of the control may evolve during rollout.
Limits and safeguards
Steering a task is not the same as rolling back completed actions. For sensitive code changes, deployments or destructive operations, teams should use version control, checkpoints, explicit approvals and test runs. New instructions can also conflict with earlier requirements, so concise corrections and clear priorities help reduce ambiguity.
NetContentSEO analysis: corrigible workflows, not just autonomous ones
The value of agentic tools is not only that they can execute longer tasks; it is also that humans can redirect work while it is in progress. For SEO engineering teams, faster follow-ups may make audits, analysis scripts and automation more responsive to discoveries made mid-run. This is a product workflow improvement, not a change to search rankings or indexing.
Official source
OpenAI Help Center — ChatGPT release notes (October 8, 2026).