GitLab Inc., the intelligent orchestration platform for DevSecOps, released GitLab 19.4. As agentic automation spreads from individual developers to entire engineering teams, what limits its reach is no longer what agents can do but how confidently an organization can extend it. GitLab 19.4 brings agentic automation to every surface developers work in, with new options for cost efficiency and the control platform owners need to scale it.
Developers can now delegate a whole objective from the terminal, run flows the moment a merge request opens, and reach GitLab from any Model Context Protocol (MCP) client through a governed set of tools. Platform owners get the tool-level controls and the per-person attribution to widen access to that automation and account for how it is used.
Both come from the platform already running the work. The same permissions that cover the code govern the agents, and every credit traces to the user account that spent it, so there is no second permission model and no separate audit trail.
AI agents have handled individual tasks well, but developers have still had to break the work into those tasks and check each result. Handing over an objective to agents also meant no independent check on their judgment.
The /goal slash command in GitLab Duo CLI automates a whole objective until it's done or reports what it could not resolve, with project guardrails and context. The agent implements the work, and a separate model verifies it against the stated goal at each step, deciding whether the goal is met or the iteration limit is hit. The developer can stop the run at any point, then revise the goal and restart, and the flow runs locally under the organization's existing rules. A developer can hand off a bounded piece of work and come back to a verified result.
No single model fits every task. GitLab Duo Agent Platform now offers three GitLab-hosted open-weight models, Kimi K3, MiniMax M3, and GLM 5.3, alongside the frontier models already available, giving teams more flexibility to choose the model best suited to each agentic automation workload. By selecting models based on task requirements, teams can balance quality, latency, and cost. The new hosted open-weight models in GitLab Duo Agent Platform get up to 4x more calls per GitLab Credit than many comparable frontier models, giving teams more model options to match cost to task complexity.
Group owners retain the same governance over model choice as every other GitLab Duo Agent Platform capability. They can set a default model for each feature and curate the models available to teams, with those settings applying across child groups and projects. GitLab evaluates each model against internal performance and quality standards, and vets every hosting vendor through its third-party risk management process, helping teams expand model choice without giving up administrative control.
