Scheduled Tasks

Run the Judgment Agent on a schedule and get its answer in Slack.

A Scheduled Task is a Judgment Agent chat that runs your instructions on its own, with the same tools and Agent Knowledge as a chat in the app. Use it for questions you’d otherwise ask every morning, like “What broke yesterday?”

Judge actions react to one evaluation at a time. A Scheduled Task looks across many traces and writes you a report.

Create a task

Open Scheduled in the sidebar. The page lists Your tasks, Suggested tasks and Built-in agents.

Select New task and pick one:

Create opens a blank task.
Template opens Templates: Daily quality review, Weekly failure summary, Investigate regressions, Review slow traces, Weekly cost review, and Find new test cases.
Suggest for me opens a Judgment Agent chat that suggests tasks for your project.

Then fill in the task:

Name this task and, optionally, Add a description.

Write the Instructions. This is the prompt the agent runs every time, so say what to look at, over what window, and what you want back. For the lunch-order agent:

Review traces from the past 24 hours in production. Find restaurants whose
get_menu calls failed, how many orders were affected, and whether the agent
told the user. Link representative traces. Keep it under 10 bullet points.

Set the Schedule, Delivery and Agent sections (below).
Select Create task.

Select Run to try it once right away (the tooltip reads Run now). Run it a couple of times and read the output before you put it on a schedule. Run uses the saved instructions, so save changes first.

Schedule

Turn on Run on schedule and choose how it Repeats:

Every day or Every week (pick the days), at a time you choose
Every 15 or 30 minutes, every hour, or every 2, 4, 6 or 12 hours
Custom cron, for example */30 * * * *

Times are in Pacific Time (PT). A task with the schedule off shows Manual only; a paused task shows Paused with its cadence.

Delivery

To get the agent’s final answer in Slack, with a link back to the full run:

Turn on Post to Slack.
Select at least one channel.

Every completed run posts, whether it started on the schedule or from Run, so turn this on after you’re happy with the output. Post to Slack doesn’t depend on Run on schedule, so you can turn it on for a task that only runs manually.

Post to Slack needs your organization’s Slack connection. Manage Slack connection takes you there, and Connections explains the setup.

Agent

Run mode: Deep Research (the default) or Fast. Both use the same tools; Deep Research takes longer and handles questions that need several queries.
Write permissions: Ask for writes (the default) or Auto-allow writes. With Ask for writes, the agent asks for approval before it changes anything, like creating a judge or adding examples to a dataset. Use Auto-allow writes only for tasks whose instructions you trust to make changes unattended.

Runs and status

Each task lists its recent runs under Recent runs; open one to read what the agent found. Cards show the latest state: Running, Needs input, Needs attention or Failed.

If a task’s schedule can’t be set up, the task shows a banner asking you to contact support.

The task sheet also has a button that asks the agent to rewrite your instructions: Tailor to my project for a task started from a template, Improve otherwise.

Who can do what

Anyone who can view the project can see tasks and their runs.
Org admins and owners can edit any custom task. A developer can edit tasks they own.
Only a task’s owner can run it, and scheduled runs run as the owner.
Built-in agents, like the one behind Issues, can’t be scheduled or deleted from this page.

Each run is a Judgment Agent run and uses Judgment Units. A Deep Research task every 15 minutes adds up; see Billing and plans.