Descriptions:
Ankit Mathur from Databricks presented at AI Dev 26 in San Francisco on the enterprise challenges of deploying coding agents at scale, arguing that AI governance — not model capability — is the primary bottleneck slowing organizational adoption. He cites industry data showing coding agent usage runs 10 to 100 times higher than all other agent categories combined, creating urgent pressure for centralized oversight frameworks.
The talk introduces Databricks’ Foundation Model API gateway as a unified access layer supporting models from Anthropic, OpenAI, Google Gemini, Meta, Mistral, and others, while enforcing privacy guarantees, cost controls, and audit logging. Mathur explains why coding agents specifically need to be the most privileged users in an enterprise — requiring access to Kubernetes logs, Jira, Slack, and internal APIs — and why most organizations struggle to grant that access safely due to data retention liabilities and compliance requirements.
Practical guidance covers monitoring power users and adoption blockers, configuring per-developer spending limits across tools, and ensuring model providers cannot retain sensitive enterprise data by default. Mathur frames the governance gap as existential competitive risk: companies that cannot roll out coding agents frictionlessly are already falling behind peers who have solved the access and oversight problem.
📺 Source: DeepLearningAI · Published May 21, 2026
🏷️ Format: Deep Dive
