What changed
Databricks made Genie Agent mode APIs generally available, allowed scalar and table-valued Unity Catalog functions to act as agent tools, and added Genie Code as a beta Lakeflow Jobs task. A separate public preview brought Genie Code into AI Runtime notebooks for distributed-training development and troubleshooting.
Why it matters
Genie moved from an interactive assistant into APIs, unattended jobs and GPU engineering workflows. That makes service identity, tool scope, upstream data, auto-approval and terminal evidence operational controls rather than user-interface details.
What to do
Start with one read-only scheduled task or API canary. Give it a dedicated least-privilege identity, remove write-capable MCP tools, cap runtime and cost, preserve the full conversation and independently verify the resulting data or ticket state.
Caveat
The Agent mode API is generally available, but Genie Code job tasks are beta and AI Runtime support is preview. Job tasks always run with auto-approval enabled, which Databricks explicitly says is not a security boundary and cannot be disabled.