A Databricks workspace took about four days to deliver, between intake and manual deployment. I built an agent for it alone and tested it in a sandbox environment before moving it to development.
- Intake. It read approved ServiceNow requests and opened fully specified Jira tickets through MCP connectors.
- Preview. Before anything ran, it re-read the ticket, checked for a duplicate workspace and printed every resource it would create, from access roles to storage and the SQL warehouse, for an engineer to sign off.
- Defaults. Workspaces it deployed started with 91 features disabled, and teams requested the ones they needed.
- Approvals. I negotiated UAT and production approvals down to one manual apply. When stakeholders kept the merge request as a gate, I reconfigured the agent to raise it and wait for approval.
- Close-out. It generated the configuration, ran plan and apply, published the release document and commented the work back onto the ticket before closing it.
- Checks. With the team, we graded each deployment against verified ground truth.
Result. Delivery went from four days to roughly six hours across 22 deployments by mid-August 2026.
Status. In September, every environment after development moved to multi-tenancy, so the agent now runs in development only.