How to use automations in ChatGPT and Codex
Set up recurring tasks in ChatGPT and Codex, test permissions, and send useful results to Agent Alerts when your workflow supports it.
A useful automation answers a question you would otherwise keep checking: did the release finish, what changed in this project, or is there anything I need to review?
Start with one question, a clear schedule, and a small result. Once that works, decide whether the result belongs in your chat, a notification, or an Agent Alerts Live Activity.
Choose where the work should run
ChatGPT scheduled tasks handle reminders, recurring updates, and monitoring. For longer work involving research, tools, and finished deliverables, ChatGPT Work provides the agent experience. Codex is the place for repository reviews, tests, and other software development work. OpenAI’s guide to Work and Codex explains the current choices.
The schedule decides when to start. The selected environment decides which files, tools, and accounts the task can actually use.
For a Codex task using a local project, keep your computer on, the desktop app running, and the project available on disk. Don’t assume a local schedule moves to the cloud when your laptop sleeps. See OpenAI’s automation documentation.
Set up a ChatGPT scheduled task
Open Scheduled, or ask ChatGPT to create a task. Give it the work, cadence, time zone, and expected output. Review the saved task before relying on it.
Try a simple reminder first:
Every weekday at 9 AM America/New_York, remind me to review my release checklist. Keep the reminder to one sentence.
You can manage tasks from Scheduled and configure push or email notifications in Settings → Notifications. Availability, active-task limits, and supported schedules depend on your account and workspace. ChatGPT scheduled tasks
For a connected workflow, name the source and keep the first version read-only:
Every Monday at 9 AM America/New_York, check open pull requests in [repository] using the connected GitHub app. Report reviews waiting on me and failing checks, with links. Do not edit files, post comments, or merge anything. If access fails, report the blocker.
Replace the repository placeholder before scheduling.
Eligible Work accounts also support event-triggered tasks for selected app events. Review the trigger and conditions. Connecting an app alone does not create that task. Required approvals can pause a run. Scheduled task triggers and permissions
Set up a Codex automation
Open the relevant project in Codex and test the prompt manually. Then ask Codex to schedule it. Specify whether future runs should continue the same chat or produce independent reports.
For example:
Every weekday at 9 AM America/New_York, review commits from the previous 24 hours in this project’s local Git history. Identify concrete regression risks with file references. Do not edit files, install dependencies, push commits, or contact anyone. If there are no actionable findings, say so briefly. Create a separate report for each run.
When changes are part of an authorized workflow, a Git worktree can separate them from unfinished local work. Scheduled runs use sandbox and permission settings; restricted operations may fail. Start with narrow access rather than enabling full access to make a failing task run. Codex scheduling and security
Review the first few runs. Check the source window, findings, and actual execution time before increasing the frequency.
Add Agent Alerts after the task works
Agent Alerts displays the update your workflow sends. Your AI client or scheduler still owns the recurring work. Connecting Agent Alerts does not start monitoring.
A send step is appropriate only when that run has a working Agent Alerts integration, the required tool, and permission to send the intended information. MCP availability can differ between clients, hosts, and plugins. A connection in one environment does not establish access everywhere. OpenAI’s MCP documentation
After testing the connection, you can add a scoped instruction:
When this check finds a blocker or completes the requested review, use the available Agent Alerts integration to update activity ID
release-review. Send only a neutral status and next step. Keep private repository details out of Lock Screen text. If the send fails or requires approval, report that in this chat.
Follow the Agent Alerts setup guide, send a non-sensitive test, and check your iPhone. An accepted request does not prove visible delivery.
Keep the automation worth running
Choose a cadence that matches the decision. A daily project review rarely needs minute-by-minute checks. A temporary release watch should have an end condition.
Write down:
- Which source the task may read.
- What counts as a meaningful change.
- What it may send or modify.
- When it should stop or ask for help.
Keep one stable activity ID for the same subject, and update it when there is something useful to show. Keeping one Live Activity current explains that pattern.
The goal is a result you can trust, with enough context to know what to do next.
Choose a workflow for your role
Start with a decision you already make. These guides show concrete checks, useful reporting windows, and the actions that still need your review: