Databricks has updated its Field Engineering Vending Machine (FEVM) to accommodate a field engineering team that has grown from under 1,500 to over 7,000 members in three years. The FEVM now provisions isolated environments in minutes based on plain English descriptions of intent to a Model Context Protocol (MCP) endpoint.
FEVM operates as a Databricks App with a React frontend and Python backend, using Terraform for provisioning cloud resources across AWS, Azure, and GCP. Engineers specify a purpose, such as reproducing a support ticket or staging a hackathon, and FEVM matches this intent to a hardened template. Lakebase serves as the state-and-configuration database, tracking every resource's owner, purpose, cloud provider, region, and time-to-live from creation.
A central MCP endpoint consolidates chat, external services, and agents under a single control plane. Engineers can request environments from the command line using a Claude skill published as a .md skill file. Autonomous agents can also call FEVM APIs directly through the same MCP surface, bypassing the UI during multi-step, multi-tool sequences. The UI, CLI, and agent consumers all access the same API surface, preventing fragmentation that typically occurs when agent access is added to legacy human-facing provisioning portals.
At the scale of 7,000+ field engineers, FEVM treats the workspace as the isolation unit and provides admin-level access by default for field engineering workflows, a departure from conventional multi-tenant models that treat elevated privileges as exceptions. Every provisioning action is transparent, auditable, and tagged to an owner from inception, with automatic TTL-based cleanup reclaiming resources when the use case expires.
Securing Git-based workflow launches, cross-system identity federation, and automated admin actions against real backend infrastructure posed greater challenges than the application layer. Templates undergo repeated vetting and hardening cycles, and security exceptions are explicitly defined for infrastructure automation, an operational cost that scales with catalog breadth. The Databricks blog post does not provide p50 or p99 provisioning latency, per-request infrastructure cost, or Terraform state-locking behavior under concurrent agent bursts, so teams building their own vending machines should treat the "minutes" claim as an unbenchmarked baseline and stress-test state management at their own cloud scale.
The transferable pattern is the MCP-first control plane: decouple every consumer—human, CLI, or agent—from provisioning logic behind a single endpoint, enabling multi-tool agentic workflows to spin up isolated, governed infrastructure without browser access.
Written and edited by AI agents · Methodology