Some workflow steps stay open for days or weeks — tracking a work item, sending reminders, or re-checking an external system on a cadence. Scheduled work lets Thunk.AI run the step's AI agent on a recurring timer while the step remains active.

Each timer tick is a normal agent run on that work item, driven by the step's AI Instructions. The agent does its work for this pass and, by default, waits for the next scheduled run rather than finishing the step. See How an AI agent ends a run for that conclude behavior.

Scheduled work is configured once per workflow step in the plan. Every active work item on that step shares the same schedule, filter, and scheduled-run instructions.

Use a schedule when a step should wake up without a new inbound message — for example:

For a full pattern that combines inbound updates with scheduled checks, see Design pattern: the Work Tracker.

Turn on a recurring schedule with Daily at specific time, Every hour, or Monthly on a specific day. When a schedule is active, an enabled indicator appears on the section header.

While authoring the workflow plan (no work item open), you can edit the schedule shape but cannot fire a test — testing requires a concrete work item and step instance.

Choose a Daily Time, Timezone, and optional Days of week chips (for example weekdays only). The agent runs once per selected day at that clock time in the chosen timezone.

Runs at the top of each hour. Optionally turn on Only run between certain hours, pick From / To hours, and set a Timezone so ticks happen only during business hours in that zone.

Choose a Timezone, a Time, and a Day of month. The agent runs once a month on that day at that clock time in the chosen timezone.

Pick a numbered day (1–31) or Last day of month. A numbered day of 29, 30, or 31 only fires in months that have that day — shorter months are skipped — so choose Last day of month when you want the step to run every month (for example, end-of-month billing or reporting).

On the first time a work item reaches the step, the agent runs once to do any setup, then typically concludes while leaving the step open.

On each timer tick, the agent runs again on that work item. Unless your instructions say to finish the step, the agent concludes by waiting for the next scheduled run — the step stays active for the next tick or for an inbound update.

Write directions that separate:

Use a Step Finish Condition under Workflow Options when the platform should enforce when the workflow may proceed. See Control when a step runs and proceeds.

By default, every active work item on the step is eligible on each tick. On thunks with many rows, that can be expensive.

Query Filter narrows scheduled runs to work items whose property values match a condition — the same condition builder used for Workflow Options and tests. You must bind input properties on the step before a filter is available.

Examples:

On a real scheduled run, work items that do not match the filter are skipped for that tick — the agent is not invoked for them.

By default, scheduled runs use the step's Directions field. Turn on Use custom AI instructions for scheduled runs to author separate directions that apply only on timer ticks.

When this override is enabled, those directions replace the step's Directions for scheduled runs. Include any phrasing about when to finish the step or keep waiting in the override — the scheduled run reads these directions to decide.

Interactive runs (for example when an inbound message arrives) still use the step's normal Directions.

From a work item on a step with an active schedule:

Both open the Scheduled Work dialog. You can:

Simulated time affects what the agent is told the current time is. Tool results, message timestamps, and data the run writes still use the real clock.