GitLab releases version 19.4 with goal-driven AI agents

GitLab has launched version 19.4 of its DevSecOps platform, introducing goal-driven AI agents and enhanced cost management features. A key addition is the /goal command for GitLab Duo CLI, currently in public testing. This feature allows agents to execute tasks autonomously based on open-ended objectives, using one model to implement changes and a second model to verify progress against the stated goal.

The update also expands the GitLab Duo Agent Platform to support hosted open-weight models, including Kimi K3, MiniMax M3, and GLM 5.3. GitLab notes these models can offer up to four times the call volume per credit compared to some frontier models, allowing enterprises to optimize costs by assigning specific tasks to more economical models.

To manage the implications of increased automation, GitLab has introduced granular cost attribution and usage controls. Administrators can now set default usage limits for users and receive alerts when consumption reaches 50% or 80% of monthly limits. Additionally, expanded Model Context Protocol (MCP) tools allow agents to interact with pipelines, merge requests, and codebases under strict permission settings, with write and delete operations defaulting to requiring manual approval.