How to Govern and Audit Your Coding Agent Usage with Databricks
Overview
| Experience | In Person |
|---|---|
| Track | Artificial Intelligence & Agents |
| Industry | Enterprise Technology |
| Technologies | Databricks Agents |
| Skill Level | Beginner |
Session Speakers
Aarushi Shah
/Software Engineer
Databricks
Rohit Agrawal
/Senior Staff Software Engineer
Databricks
Full Summary
Coding agents at scale: developer freedom with admin governance
Coding agents are surging inside enterprises, often dwarfing other agentic use cases in token consumption. Teams want the productivity lift, yet the rapid churn in models and harnesses, plus fragmented security and billing, creates friction for developers and headaches for admins. The talk introduces Omnigent, an open source "meta harness," and shows how pairing it with the Unity AI Gateway for enterprise AI governance gives engineers choice without sacrificing control. For a broader look at how AI is reshaping software delivery, responsible AI practices on the Databricks Platform illustrate the governance principles at work here.
FAQ
Yes. Omnigent is presented as an open source developer tool available at omnigen.ai and via the Databricks OSS GitHub repository. Developers can download it and run a one-time omnigent configure to pull the admin-managed policy, with the full admin experience designed to pair with Unity AI Gateway.
No. Omnigent launches each harness in its native UI, so commands such as omnigent claude or omnigent codex drop you straight into those tools. Authentication, MCP setup, and policy enforcement happen in the background.
Behavior is policy driven. Admins can warn at thresholds, block further usage at a cap, or automatically fall back to a cheaper model and harness. The demo showed warnings around 80 percent and a hard stop at 100 percent of a daily budget.
Admins define security hooks that inspect prompts before model calls. During the demo, a pasted Databricks authentication token was automatically flagged and rejected, a pattern that can extend to other PII or secrets the organization chooses to block.
Yes. Admins can approve specific MCP servers such as GitHub, Slack, Atlassian, and Databricks SQL, then control which individual tools within those MCPs are exposed. For example, they can grant read-only access while withholding write actions, applied uniformly across connected coding agents.