GitLab 19.4 Adds Agentic AI Controls and Cost Management

GitLab expands AI agent capabilities with enterprise controls

GitLab has shipped version 19.4 of its DevSecOps platform with expanded agentic automation features designed to move AI assistance from individual developer tools to organization-wide deployment.

The release centers on a new /goal command in GitLab Duo CLI, now in public beta. Instead of directing AI through discrete steps, developers can assign an open-ended objective from the terminal. The agent runs locally, monitors its own progress, and uses a separate verification model to confirm whether it has met the goal or hit iteration limits. Developers retain the ability to halt execution, adjust the objective, and restart.

According to Manav Khurana, GitLab's Chief Product and Marketing Officer, the shift represents a move from task-based assistance to bounded delegation. "The platform running the automation is what governs which tools an agent can touch and attributes what it consumes, so extending it to the next team is a measured decision rather than an open-ended risk," Khurana said.

Why it matters

As enterprises move AI tools beyond pilot projects, finance and platform teams face mounting pressure to track spending and establish accountability. GitLab's approach ties agent permissions to existing code and workflow governance, avoiding the need for separate AI-specific control frameworks—a practical concern for organizations deploying agents across multiple engineering teams.

Multi-model support targets cost optimization

GitLab 19.4 adds three hosted open-weight models to its Duo Agent Platform: Kimi K3, MiniMax M3, and GLM 5.3. The company states these models can deliver up to four times more API calls per GitLab Credit than comparable frontier models, giving teams the option to match model selection to task requirements based on quality, speed, and budget.

Group owners control which models teams can access. Administrators can set default models for specific features and restrict available options across groups and projects, with settings cascading through child structures.

Cost visibility is now generally available, providing platform owners with a dedicated settings interface for per-user spending caps and exports at the billable event level. Individual developers can also view their own consumption.

Model Context Protocol integration extends agent reach

The release includes public beta support for Model Context Protocol server tools, allowing agents operating from external clients to execute actions within GitLab workflows. Supported operations span CI/CD pipeline triggers, merge request management, work item updates, vulnerability triage, and project administration.

These external agent actions remain subject to the same group and project permissions applied to GitLab Duo Agent Platform. Read-only tools default to Always Allow status, while write and delete operations default to Always Ask.

Additional features

GitLab 19.4 also introduces an experimental Slack integration for Premium and Ultimate customers, enabling users to query projects, create issues, and retrieve context-aware information through threaded conversations. Model selection for the Developer Flow is now generally available as a standalone configuration option.

Community contributors expanded MCP server functionality with tools for reading project metadata, listing repository branches and merge requests, and retrieving previous Duo sessions.

These details were first reported by IT Brief Australia.