Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding

Export Jules work and rebuild scheduled tasks before the shutdown

Written October 31, 2026

For anyone still running work in Google's Jules coding agent. Jules winds down at the end of the year, and the useful deadline is earlier: after 11 November you can look but not work. Export first, check the export, and only then rebuild each recurring task where it will live next. Nothing in this move happens on its own.

The dates that matter

Google's Jules turndown notice says Jules will wind down on December 31, 2026. You have until November 11, 2026 to keep finishing sessions. After November 11, Jules turns read-only until the end of the year. Read-only means the history is still there to look at, but no new work runs. Do the export before November 11, while sessions can still finish and while you can still fix a gap the export reveals.

Your code is not the thing at risk. The notice says your code, branches, and pull requests already live in GitHub, so they stay where they are. The export covers the configuration and history that live inside Jules: settings, environment variables, scheduled tasks, connected repositories, memories, setup scripts, and sessions. That is the part with no other copy.

Run the export

The notice gives one export path. Follow it in this order:

  1. Install Jules Tools and sign in. Run the two commands from the notice:
    npm install -g @google/jules
    
    # then sign in:
    jules login
  2. Connect GitHub first, if you use Jules with GitHub. The notice says to make sure your GitHub account is connected before exporting, or the export will not include repository data. If you have never used Jules with GitHub repositories, the export will simply have no repository data.
  3. Open a command-line coding agent from your home directory and paste in the export prompt from the notice. The notice names Claude Code as an example of such an agent. The prompt tells the agent to check the setup, run the export, and report what it saved: the counts it prints, and the full path of the folder. The prompt also tells the agent not to print your access token or the values in the environment variables file. Use the prompt as the notice prints it, rather than paraphrasing it from memory.

If the agent reports an error from one of the calls the export makes, stop and read what it says before trying again. The notice's prompt warns that the calls it uses are undocumented, and that a call returning an error should be reported as-is rather than worked around by guessing. A guessed fix can silently produce a folder that looks complete and is not.

Check the export folder

The export lands in ~/jules-export. Do not assume it is complete because the folder exists. Open it and match what is inside against the list in the notice:

Open two or three session folders all the way through: summary, activity log, and at least one diff. A summary with no diff behind it, for a session you know produced code, is a gap to chase while Jules can still run. Also open the scheduled tasks file and read each task's prompt and schedule. That file is the raw material for the rebuild below — if a task is missing from it, it will not be rebuilt from memory later.

Keep the export folder private and out of your repositories. The notice says the environment variables file contains your environment variable values. Treat the whole folder the way you would treat a file of secrets: no commits, no pastes into chat logs, no shared drives.

Rebuild each scheduled task by hand

There is no automatic move. The notice offers an export to a local folder. It does not move a task, a memory, or a setup script into another product for you. Each scheduled task has to be recreated where it will run next, one at a time, from the prompt and schedule in the export.

Start from the scheduled tasks file you just checked. For each task, write down three things before you touch the new tool: the task's prompt, its frequency and cadence, and the repository it ran against, with that repository's setup script and memories beside it. In Jules, a scheduled task was made from the task input field through the Planning dropdown, with its frequency and cadence set at the top of the input field, as Jules's scheduled tasks documentation describes. Editing a scheduled task was not supported there either — the documented way to change one was to delete it and create a new one — so rebuilding from the prompt and schedule is the same shape of work Jules itself required.

If your next home is Google Antigravity, the destination is Automations. The Antigravity changelog says the Scheduled Tasks tab is now called Automations, where you can run an automation on demand or on a schedule, or set one up with the agent's help. For each Jules task, use Create with Prompt and let the agent walk through the setup, giving it the prompt and the cadence from the export. Then run the new automation once with Run Now and read what it did, before trusting it to run on its own. A task that ran in Jules against a remote machine with a setup script needs its setup carried over in the new automation's instructions — the export's setup script and memories are the source for that, not your memory of what the task did.

Do not wait for a closer match to appear. The turndown notice says what Google learned from Jules is shaping Antigravity, and it gives no date for background execution of the Jules kind inside Antigravity. Build on what Antigravity offers now, and re-check its changelog if the feature you need arrives later. Waiting past November 11 for a feature with no date leaves you with a read-only Jules and no export left to fix.

The trap: a schedule that lives in one conversation

Antigravity has two different scheduling mechanisms, and mixing them up loses work quietly. The /schedule command sets a timer or a recurring schedule that wakes the agent in the current conversation. The Antigravity schedule command documentation states the scope plainly: timers and schedules belong to the conversation that created them, and for standing jobs that outlive a session, use a persistent sidecar or Automation instead. Its comparison table puts it in one line: /schedule lasts for the current conversation, while Automations and sidecars persist across sessions and restarts.

That makes /schedule the right tool for work inside a task you are already running — checking whether a deployment finished, polling a test run, a follow-up later in the same conversation. It is the wrong tool for a rebuilt Jules task that must keep running when no conversation is open. A recurring job rebuilt with /schedule inside a chat session does not become a standing job. Rebuild recurring Jules tasks as Automations, and reserve /schedule for timers and follow-ups inside live work.

What does not move on its own

Limits

Related playbooks

Sources

Sources: Jules documentation, including the Jules Turndown notice, Jules scheduled tasks documentation, Antigravity changelog, Antigravity schedule command documentation.

Back to: Coding playbooks hub