How do we roll out coding assistants to employees?

Summary

• Roll out in stages and decide the controls before you widen access: approved assistants and models, how credentials are issued, which tools an assistant may use, what data may leave, and who sees usage and spend.
• Run a contained pilot with a small group of interested teams, trained champions, success measures defined in advance, and one place for questions. Use what the pilot reports to refine policy before each wave.
• Route model traffic through one governed gateway. With Unity Gateway, developers use coding agents such as Claude Code and Codex without provider API keys on their machines, while permissions, rate limits, budgets, and usage tracking apply centrally.
• Measure adoption, engagement, and satisfaction, not only usage counts. Databricks combines system table metrics with a developer survey and treats productivity gains and time saved as self-reported.
• Keep human review: generated code goes through the same review and CI as any other change.

How do we roll out coding assistants to employees?

Roll out in stages, and decide the controls before you widen access. Settle policy first: which assistants and models are approved, how credentials are issued, and what data may leave. Then run a pilot with a small group that reports back through one channel, and expand in waves while you track adoption and outcomes. Routing model traffic through one governed gateway turns the policy into something you can enforce and audit instead of a document.

What should you decide before anyone gets access?

Decision	What to settle	How it is enforced
Approved assistants and models	Which tools and models employees may use	EXECUTE on Unity Gateway model services; assistant-side settings such as Claude Code's availableModels
Credentials	Who holds provider keys	The Unity Gateway CLI (ug) signs developers in with Databricks credentials, with no provider API key on their machines
Permissions	Which tools, commands, and network destinations an assistant may use	Claude Code managed settings: permission rules and sandboxing
Tool access	Which MCP servers agents may reach	MCP servers registered in Unity Catalog, with tool filtering
Spend and capacity	Per-user limits	Rate limits per user (requests or tokens per minute) and budgets that alert or block
Sensitive data	What may reach a model	Service policies (Beta), such as the built-in sensitive data detection policy
Audit	What you can review later	The system.ai_gateway.usage table; inference tables for request and response payloads

How do you run the pilot and expand?

Start with a small group of interested teams. Define success measures and train champions before launch, give questions one place to land, and hand each pilot user one real task to try. Ask what worked, what was annoying, and what surprised them; that feedback decides how you roll out to everyone else. Then expand in waves, adjusting enablement and policy after each one. Admins can publish a workspace configuration: ug configure applies the admin's settings, and plain ug launches the agent the admin selected, so every wave starts from the same setup.
Keep review in place. Generated code goes through the same code review and CI as handwritten code. Genie Code is governed by Unity Catalog permissions, so it can only reach the data and operations that the user can. Databricks recommends always reviewing code that Genie Code generates before running it. In Agent mode, Genie Code asks for approval before it runs code or edits a notebook, but the default mode for new users is Auto-approve. Databricks calls that a productivity feature, not a security boundary, and advises keeping it off when you work with production data.

How do you measure whether it is working?

Measure three things: adoption (top users, submission counts), engagement (active users, top tasks, frequency of use), and user satisfaction (productivity gains, time saved). The Databricks guide for Genie Code combines system table data with an internal survey and offers a downloadable events dashboard. It treats productivity gains and time saved as self-reported survey measures.
For external coding agents routed through Unity Gateway, the built-in usage dashboard has a Coding Agents tab. It tracks integrated agents such as Claude Code, Codex, and Cursor with active days, coding sessions, commits, and lines of code added or removed.

Which assistants can Unity Gateway govern?

Unity Gateway connects coding agents such as Claude Code, Codex, and OpenCode to models, MCP tools, and shared skills. Coverage varies by agent. For Cursor Agent, ug cursor configures MCP access only, and model requests use your Cursor account. Check the capability matrix in the Databricks docs before you promise a control to your security team.

FAQs

Should developers bring their own API keys?

No. For agents that route through Unity Gateway, every request goes through a model service using Databricks credentials. EXECUTE permissions, rate limits, and budgets then apply to that model service, and usage is recorded centrally.

How do we keep sensitive data away from a model?

Attach a service policy to the model service. The built-in sensitive data detection policy looks for configured data types, such as credit card numbers, in requests and responses and blocks or redacts matches. Service policies are in Beta, and an account admin must turn on the Unity Gateway beta features first.

What should the pilot decide?

Whether adoption, engagement, and satisfaction met the measures you set before launch, which controls caused friction, and what to change before the next wave. Settle those answers before you expand.

Can we restrict which MCP servers assistants use?

Yes. Unity Gateway governs MCP servers registered in Unity Catalog with grants and tool filtering. Claude Code also offers managed settings that allow or deny MCP servers for your organization.