Skip to content
EN

Scheduled tasks and follow-ups

Open the scheduler and choose a project. Create a task with an instruction, a once, daily, weekly or cron schedule, an IANA timezone and a missed-slot policy: skip or run the latest missed slot once.

Choose New work to create a Thread for each occurrence. Select a registered model, source and workspace access: read-only snapshot, isolated editable copy or live workspace. You can add a Continuing Goal with an optional output-token budget. Existing work instead targets an exact Thread and branch and inherits its model, source, permissions and goal.

Enable scheduled fires when ready. Run now accepts an occurrence; it does not mean the work completed. Open its conversation for results and controls. Disabling a definition holds undelivered scheduled work; stopping a run does not disable future slots. The Application Host must be running.

Project .agents/loops/ and user ~/.agents/loops/ files can define cron tasks. Replace the model IDs below with IDs from your saved model channel:

---
name: daily-review
runtime: agent
schedule: "0 9 * * *"
timezone: UTC
enabled: false
target:
kind: new_work
model: { providerId: my-channel, modelId: my-model }
source: { kind: project }
sourceMode: fixed_branch
goal: null
---
Review repository changes since yesterday.

The file owns the definition. Edits use revision checks; discovery does not rewrite files. runtime: agent, a native target, an explicit timezone, name, cron schedule and body are required. Unsupported old definitions are not automatically imported or rewritten. Invalid replacements hold the last valid native task until repaired.

Use the scheduler’s Follow-ups view or Schedule a follow-up inside a conversation. A follow-up continues that conversation after a time, process/job stop, file condition, research artifact/collection, log event or machine-metric threshold. Conditions can be combined as any or all.

Choose once, or each observed change for supported repeatable conditions. Pauses, unanswered questions and goal limits remain in effect. Observation gaps can miss changes. Pause, resume or cancel the registration and inspect its occurrence history. Cancelling prevents future and undelivered work; existing results and observed jobs remain.