GitLab 19.4 adds agentic automation with cost controls

GitLab Inc. has shipped version 19.4 of its DevSecOps platform with a focus on autonomous agent capabilities that balance automation power with enterprise cost management.

The centerpiece is a new /goal slash command in GitLab Duo CLI that automates entire objectives rather than individual tasks. The system uses a dual-model architecture: one agent implements the work while a separate verification model checks progress against the stated goal at each step. Developers retain control throughout, with the ability to halt execution, revise objectives, and restart under existing organizational policies. All processing runs locally within the company's governance framework.

Expanded model selection for cost optimization

GitLab Duo Agent Platform now hosts three open-weight models—Kimi K3, MiniMax M3, and GLM 5.3—alongside existing frontier models. According to GitLab, these new options deliver up to four times more API calls per GitLab Credit compared to many frontier alternatives, letting teams match model selection to task complexity and budget constraints.

Group owners control model access through the same governance interface used for other platform capabilities. They can set default models per feature and curate which models teams can access, with settings cascading to child groups and projects. GitLab vets all models against internal benchmarks and puts hosting vendors through its third-party risk management process.

New automation tools and governance controls

The GitLab Model Context Protocol (MCP) server gained tools for triggering pipelines, reading failed job traces, managing merge requests from creation through merge, searching work items, and triaging vulnerabilities. Administrators govern these through existing GitLab Duo Agent Platform rules configured at the group and project level.

Read-only tools default to "Always Allow" for uninterrupted lookups, while write and delete operations default to "Always Ask," creating approval checkpoints before agents modify anything.

Usage visibility and cost management

Platform owners now have granular visibility into GitLab Credits consumption. Per-user caps appear on a dedicated settings page, and usage exports drill down to individual billable events. These reports arrive via email with secure download links for both GitLab Flex and standard subscriptions. Developers can view their own consumption for the first time, enabling team-level budget management.

Additional capabilities

GitLab Duo Agent Platform integration with Slack entered experimental availability for Premium and Ultimate customers. Teams can mention @GitLab in threads to search the platform, open issues, or query projects using conversation context. The agentic workflow executes on a CI/CD runner and posts results back to the thread.

Model selection for the Developer Flow in Duo Agent Platform reached general availability, letting administrators choose models independently for different workflows. A redesigned session details panel surfaces status, timestamps, and triggering users while separating what initiated an agent session from its output.

Community contributors through GitLab's Co-Create program added MCP server tools for reading project metadata, listing branches and merge requests, and fetching previous Duo sessions. They also improved merge request status explanations and added preview rendering for Markdown and AsciiDoc files before initial commit.

Why it matters

As enterprises deploy AI agents across development workflows, cost predictability becomes a blocking concern. GitLab's combination of open-weight model options, granular usage controls, and per-user caps addresses the economic uncertainty that has slowed agentic automation adoption in regulated industries. The governance-first approach—with write operations requiring explicit approval—gives platform teams a path to enable autonomous workflows without sacrificing compliance or budget control.

These details were first reported by Automation Watch.