SoftTech AI Registry
Databricks · Genie Code

Scheduled tasks

Genie Code scheduled tasks let a user save a Genie Code prompt together with a recurring time schedule. At each scheduled time Genie Code executes the stored prompt as a full background Genie Code session without the user present, and each run produces a new Genie Code chat that can afterwards be inspected or continued interactively. Distinct from interactive Agent mode: initiation is time-scheduled rather than human-prompted, and Auto-approve is always enabled and cannot be turned off for scheduled runs.

Recorded characteristics

Function
A user saves a Genie Code prompt together with a recurring schedule; Genie Code then runs that prompt in the background at the scheduled times without the user being present. Creation mechanisms documented by Databricks: - describe it in chat (ask Genie Code in natural language to run something on a schedule; Genie Code may ask clarifying questions and then creates the task); - start from a suggested scheduled task in the Schedules pane of full-page Genie Code (also reachable from the kebab menu in the side pane); - create manually from the Scheduled tasks pane using Create > Create manually. Configurable fields documented: Title, Prompt, Schedule, Timezone. Management operations documented: Run now (manual trigger), pause by turning the task off, Edit (title, prompt, schedule, timezone), Delete, and Run history (each run opens as a Genie Code chat). Execution model: each scheduled run executes the stored prompt with the same capabilities Genie Code has in an interactive session, and each run creates a NEW chat rather than appending to one ongoing conversation. The resulting chat appears in the chat list and can be opened to review what Genie Code did and continued interactively. Distinction from interactive Agent mode: Agent mode is initiated by a human prompt in an interactive session with the user present, configurable approval behaviour and in-session interaction; scheduled tasks are initiated by a stored time schedule, run unattended in the background, and run with Auto-approve always enabled and non-disableable. Initiation is time-scheduled, unattended and autonomous after configuration; it is not event-triggered. Run now provides an additional manual execution mechanism alongside normal scheduled initiation. Explicitly excluded: Genie Code task for Lakeflow Jobs (a separate Beta capability with job orchestration, job parameters, retries and job settings) and other Databricks scheduling mechanisms.
Data access
Each run uses the stored task prompt and executes with the permissions of the user who created the task, accessing Databricks data and objects available to that user at execution time, governed by that user's Unity Catalog permissions. No run-to-run conversation carry-forward is established: Databricks documents that every run produces its own chat rather than adding to a single ongoing conversation. Previous scheduled-run chats are not documented as being supplied as context to later runs, and no such persistence is asserted here.
Actions
Can take actions
External actions
No
Human confirmation
Not required
Permission basis
User permissions
Administrative control
Documented controls are primarily user/task-level rather than a dedicated administrator kill switch: - creating, editing, pausing (turning off), deleting and manually running (Run now) a scheduled task; - reviewing prior executions through Run history; - Genie Code agentic prerequisites, including the partner-powered AI features settings that Genie Code's agentic capability depends on; - the creating user's Unity Catalog and other applicable Databricks permissions, which bound what a scheduled run can access or do; - Genie usage limits and Genie budgets/cost controls; - applicable Geo availability requirements for Genie Code features and compliance security profile workspace restrictions. No Scheduled-tasks-specific administrator-only enable/disable control is documented separately from these; no such control is asserted here.
Default state
Conditional
Availability
Generally available since 1 September 2026; introduced in Beta on 23 July 2026. The current Databricks documentation page carries no Beta or Preview label. Availability depends on the workspace having access to Genie Code's agentic capabilities, including the partner-powered AI prerequisites and applicable Geo availability for Genie Code features. Current documentation indicates scheduled tasks are not available by default for workspaces with the compliance security profile enabled. Announced future change (not yet current): Databricks states in "What's coming?" that Genie Code scheduled tasks will be available by default for workspaces with the compliance security profile enabled in October 2026. This is recorded as announced future behaviour only and is monitoring-relevant. No exhaustive region/availability matrix is published for this feature; none is asserted here.
Licensing
Scheduled runs count toward Genie usage. Databricks documents that if the Genie Code usage limit is reached, Genie Code still attempts the scheduled run and it fails in the same way an interactive session would. Genie budgets and cost controls apply. No guaranteed per-run monetary cost is documented or asserted.
External model or provider
Not fixed. Scheduled runs execute as full Genie Code sessions and therefore rely on the same Genie Code model/provider mechanism, which depends on partner-powered AI features; Databricks describes provider infrastructure dynamically rather than naming a single permanent model or provider for this feature. Where partner-powered AI features are disabled, Genie Code remains available without its agentic capabilities. No scheduled-task-specific model selection control is documented.
Limitations and uncertainty
- Scheduled runs occur without the user present to approve individual tool actions. - Auto-approve is always on for scheduled runs and cannot be turned off for a scheduled run. - Auto-approve uses an AI classifier that reviews each proposed action against the task prompt and blocks actions falling outside its scope; Databricks describes this as a best-effort productivity mechanism and explicitly not a security boundary, able to approve unsafe actions or block safe ones. An AI classifier is not human confirmation. - What a scheduled run can do remains constrained by the creating user's permissions. - Each execution creates a separate chat; no conversation history carry-forward between runs is established. - Scheduled tasks themselves cannot be shared with other users, although individual chats produced by scheduled runs can be shared like any other Genie Code chat. - Destructive-action boundary: the Scheduled tasks page establishes only that runs have the same capabilities as an interactive session and does not independently enumerate destructive operations; permitted writes and explicitly requested destructive operations are documented for interactive Agent mode. No separately documented destructive scheduled-task examples exist, and none are claimed here. - Unresolved: behaviour if the creating user later loses access; behaviour if the creating user is disabled or deleted; whether service-principal ownership of a scheduled task is supported; precise run-time permission re-evaluation semantics; whether configured MCP connections are inherited by scheduled runs; scheduled-task-specific audit logging. - The announced October 2026 compliance-security-profile default availability has not yet taken effect.

Evidence