Schedule a weekly report¶

Build an agent that writes a short report from the data in its task and publishes it as an artifact under a stable link. Then put it on a weekly schedule. The fixture includes a row that cannot be used, so you can check that the agent reports it instead of hiding it. This is a procedure to run, with one recorded run as a reference.

The page separates two questions. Does the report come out right? Test that with Run now. Does the schedule deliver it? Only a scheduled fire answers that.

Build the agent¶

Use a running installation with a model profile. No sandbox and no embedding model are needed.

Create an agent in Agents → New agent and select your model profile.
In Toolbox, enable Charts and Artifacts.
Set a budget and a step limit for the trial. The recorded runs used 15 steps and cost about 0.05 USD each.
Set the instructions below, then Publish.

You write short reports from data given in the task.
Use only the rows in the task. Report totals per category and name any row you could not use.
Label the report as synthetic when the task says the data is synthetic.
Publish the finished report with publish_artifact under the name weekly-report.

The artifact's name is its identity. Every run of this agent that publishes weekly-report updates the same artifact, so the link you share stays the same. A run adds a version only when the page changed. Identical content answers unchanged and keeps the latest version.

Create the schedule¶

Open Routines → New schedule, or the agent's Availability tab, and choose the agent. Put the data in the message, so a later run does not depend on a file somebody uploaded to an earlier chat:

Create a report for the synthetic period Demo Week.
Use only these CSV rows:
category,amount
Supplies,20
Supplies,30
Travel,15
Travel,abc
Report totals by category in a fictional demo currency and draw a bar chart.
Label the report synthetic and name any row you could not use.

For the cadence, choose At a set time, then Days of the week, tick Mon and set Time (UTC) to 09:00. The equivalent cron expression is 0 9 * * 1. The scheduler works in UTC, so convert from your local time. Save, then check the next fire time the schedule shows.

The schedule runs as the member who created it, with that member's access, and its runs are billed to the agent's budget like any other. Concepts explains why.

Run it now¶

Press Run now on the schedule. It fires one extra time and leaves the weekly cadence unchanged. The request returns as soon as the worker accepts the fire. The run then appears in the schedule's own conversation, under Routines in the chat sidebar.

Check	Reference
Totals	Supplies 50, Travel 15
The row `Travel,abc`	Named as unusable, and not counted
Label	The report says the data is synthetic
Artifact	Artifacts lists `weekly-report`, private to you
The run in Activity	Surface `schedule`, status completed
Run now a second time	The same artifact and link: a new version if the page changed, or `unchanged` if the content is identical

Open the artifact page and read the report there, not only the chat reply. The page is what people will open. It stays private until you share it or create a public link.

Recorded on v0.0.504, 25 September 2026

Model: Claude Sonnet 4.6 through OpenRouter. The schedule reported its next fire as Monday 28 September, 09:00 UTC. Run now was accepted with 202, and the run completed on the schedule surface about 30 seconds later, for 0.044 USD.

The agent called publish_artifact with the name weekly-report and got version 1, private. It then called create_chart. The artifact showed Supplies 50 and Travel 15, a synthetic-data label and "Rows excluded (could not be used): Travel, abc". A second Run now added version 2 to the same artifact and link.

The chart appeared in the run's conversation, not in the artifact. The artifact page has no network access, and the agent wrote the report without an embedded chart.

What the schedule does not decide for you¶

Where the data comes from. This fixture is fixed in the message. A real report needs a source the agent can reach on every run, such as a knowledge collection, an MCP connection or a sandbox workspace. A schedule cannot guess which newly uploaded file replaces last week's.
The reporting period. Name it in the message or have the agent read the date. A file called "weekly" does not tell the model which dates to include.
Who reads it. A new artifact is private to the person the run was for. Share it, or create a public link, on the artifact page. Artifacts covers visibility and grants.
Approvals. Neither tool used here needs one. If you add a tool that does, a scheduled run parks until somebody decides, so name who watches the approvals queue.

Watch a real fire¶

Run now proves the task, not the schedule. After the first Monday at 09:00 UTC, check that a new run appeared in Activity on its own, that the artifact gained a version and that the people who should read it can open it. The Routines dashboard card shows each routine's last outcome, so a schedule that starts failing is visible.

A schedule whose creator can no longer run the agent disables itself and records why. See concepts.

When it goes wrong¶

Nothing happens on Run now. The schedule is paused, or the background worker is not running. The worker executes every scheduled and Run now fire.
The report has no artifact. Check that Artifacts is enabled and that the instructions name weekly-report. The run's tool calls in Activity show whether publish_artifact was called and what it returned.
Each run makes a new artifact. The name changed between runs. Keep it fixed in the instructions.
A total is wrong or the bad row disappeared. Tighten the instructions before you schedule anything. A schedule repeats a mistake every week.

Record the trial¶

Keep the message, the agent version, the model profile, the cron expression, each run in Activity and each artifact version. Record which fires were Run now and which were scheduled. A person checks the totals, decides who may read the artifact and watches the first real fire.