GitLab has released version 19.4 of its DevSecOps platform, expanding agentic automation features for software development teams.
The update focuses on new controls for how artificial intelligence agents operate across GitLab, along with measures to give administrators more oversight of usage and cost. It also adds lower-cost model options and broader ways for agents to act within development workflows.
At the core of the release is a new /goal command in GitLab Duo CLI, now in public beta. The feature lets developers assign an open-ended objective to an agentic workflow from the terminal instead of directing each step manually.
The workflow runs locally and checks its own progress against the stated goal. A separate model verifies each step and determines whether the objective has been met or the iteration limit reached. Developers can also stop the process, revise the goal, and restart the run.
GitLab is positioning this as a shift from task-based assistance to broader delegation of work. Instead of using AI tools to complete one instruction at a time, developers can hand over a bounded piece of work and return to a checked result.
"This release takes agentic automation from something individual developers use to something an organisation can scale at speed and under the controls already in place," said Manav Khurana, Chief Product and Marketing Officer at GitLab.
"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.
Cost controls
Another part of the release is a wider choice of AI models in GitLab Duo Agent Platform. GitLab has added three hosted open-weight models - Kimi K3, MiniMax M3, and GLM 5.3 - alongside the larger model options already available.
According to GitLab, the new models can provide up to four times more calls per GitLab Credit than many comparable frontier models. That gives teams a way to choose different models for different tasks based on quality, speed, and cost.
Group owners retain control over which models teams can use. Administrators can set a default model for each feature and limit the list of models available across groups and projects. Those settings apply through child groups and projects, suggesting GitLab is aiming the release at larger engineering organisations with central platform teams.
Cost management is also more explicit in this version. GitLab Credits usage visibility is now generally available, giving platform owners a dedicated settings page for per-user caps and exports down to the billable event. Developers can also see their own consumption.
That matters as companies move AI tools from individual experiments to broader deployment across engineering departments. Finance and platform teams have faced pressure to understand where spending occurs and which teams or users are responsible.
Broader reach
GitLab has also expanded support for the Model Context Protocol, or MCP, through new server tools now in public beta. These tools allow agents working from clients outside GitLab to take actions across GitLab workflows.
The scope includes CI/CD, merge requests, work items, vulnerabilities, and projects. An agent can trigger a pipeline, inspect the trace of a failed job, manage merge request activity, search and update work items, and help with vulnerability triage.
These actions remain governed by the same group and project settings already used for GitLab Duo Agent Platform. Read-only tools default to Always Allow, while write and delete actions default to Always Ask.
That design reflects a broader pattern in enterprise software, where suppliers are trying to avoid forcing customers to create separate governance models for AI tools. GitLab's approach keeps agent permissions tied to the same permissions and audit trail used for the code and workflows already running on the platform.
Additional changes
Version 19.4 also includes several other additions. GitLab Duo Agent Platform in Slack is available as an experiment for Premium and Ultimate customers, allowing users to mention GitLab in a thread to search projects, open an issue, or request information tied to recent conversation context.
Model selection for the Developer Flow in Duo Agent Platform is now generally available, meaning administrators can choose a model for that flow separately from other GitLab Duo features. GitLab has also redesigned its session details panel to show status, timestamps, and the triggering user more clearly.
Community contributors added more MCP server tools, including functions to read project metadata and membership, list repository branches, list merge requests across a group, and fetch a previous GitLab Duo session. Other changes include a merge request widget that explains status in plain language and a preview showing how Markdown or AsciiDoc files will render before the first commit.
The update shows how GitLab is trying to place AI agents deeper into the software development process while giving platform teams tighter control over cost, access, and accountability at the user level.