Model choice becomes governance

GitHub and Databricks both moved this week to make model choice look less like shopping and more like operations. (github.blog) (databricks.com) On April 14, GitHub said developers can now pick a model when they launch Claude or Codex coding agents on github.com, using Anthropic models for Claude and OpenAI models for Codex inside the same workflow as Copilot cloud agent. GitHub said the feature is available “just like Copilot cloud agent,” which already lets users choose a model for a task. (github.blog) (docs.github.com) GitHub had already been widening access to those agents. On February 26, the company said Claude and Codex became available to Copilot Business and Copilot Pro customers after earlier access for Copilot Enterprise and Pro+ users. (github.blog 1) (github.blog 2) Databricks made the same argument from the enterprise side. Its April 14 launch post said Agent Bricks combines model access, execution, governance, and context so companies can build and run agents on business data “end to end.” (databricks.com) In plain terms, a model is the engine, and routing is the dispatcher that decides which engine gets which job. GitHub’s own Copilot documentation says different models trade off speed, hallucination risk, and task performance, which turns model selection into a day-to-day product setting rather than a one-time procurement decision. (docs.github.com 1) (docs.github.com 2) Databricks is adding policy to that dispatcher. The company has tied Agent Bricks to Unity Catalog and AI Gateway, which it says handle access control, auditing, centralized model access management, and failover across models and tools. (databricks.com 1) (databricks.com 2) That changes where the product sits in the stack. GitHub is packaging Claude and Codex inside existing Copilot subscriptions and interfaces, while Databricks is packaging multiple model families behind governance and orchestration controls for enterprise deployments. (github.blog) (github.blog) (databricks.com) Databricks has been building toward this for months. In March, it described Agent Bricks, Databricks Apps, and Databricks One as the path to move agents from prototype to production, and in February it said its Supervisor Agent was generally available for orchestrating multiple enterprise agents from one entry point. (databricks.com) (databricks.com) The practical result is that “which model should handle this task” is becoming a governed setting inside the platform, alongside permissions, filters, and logs. GitHub is exposing that choice to developers in Copilot, and Databricks is wrapping it in enterprise controls before agents touch company data. (docs.github.com) (docs.github.com) (databricks.com)