GitLab has introduced a goal-driven mode that allows its AI agents to tackle open-ended development tasks without requiring supervision at every step, according to the company's announcement of version 19.4 released on September 17, 2026. The update adds a /goal command to GitLab Duo CLI that accepts an objective in natural language and continues working until it finishes the task or hits an iteration limit. The platform now moves its agents beyond responding to individual prompts and predefined workflows, while maintaining permissions and approval rules already in place for each project.
The /goal command, now in public beta for Premium and Ultimate customers, uses one model to make changes and a second model to evaluate whether the result meets the stated objective. Developers can halt the process, modify the goal, or steer the agent in a different direction as the work runs locally within the project's existing context and controls. GitLab has also added support for three hosted open-weight models—Kimi K3, MiniMax M3, and GLM 5.3—which the company says deliver up to four times as many calls per GitLab Credit compared to many comparable frontier models. Group owners can now designate a default model for a feature and limit which models teams can access, allowing organizations to deploy a lower-cost model for routine tasks and save more expensive models for situations where they deliver better performance. The platform's expanded Model Context Protocol server tools now let agents in third-party clients trigger and inspect pipelines, open and review merge requests, update work items, browse repositories, and manage vulnerabilities, with most MCP tools available on Free, Premium, and Ultimate tiers.
According to Manav Khurana at GitLab, "This release takes agentic automation from something individual developers use to something an organization can scale at speed and under the controls already in place." The platform running the automation governs which tools an agent can access and tracks what resources it consumes, Khurana noted. GitLab's use of "verified" in this context has a specific meaning: the second model judges whether the outcome satisfies the natural-language goal, but this isn't formal verification of the code itself. DevOps consultancy and GitLab partner ReleaseTEAM warns that "the practical risk is automation," since a workflow connected to every pipeline failure or new merge request could execute hundreds of times daily, and recommends setting usage caps before attaching a flow to a high-volume trigger.
The release addresses cost management concerns that become more pressing when agents run automatically rather than in response to individual prompts. Credit usage exports now provide a row for each billable event, including the user, project, session, flow type, credits, and token counts, giving administrators a granular view of spending. A new settings page allows owners to establish a default per-user cap and add individual overrides without relying on the GraphQL API, while developers can monitor their own consumption and billing managers receive alerts when on-demand use reaches 50% or 80% of a monthly limit. Read-only MCP tools default to "Always Allow," while write and delete operations default to "Always Ask," with administrators able to adjust those rules at the group or project level. The /goal command can report work it couldn't resolve, and the iteration limit prevents it from running indefinitely, meaning clear acceptance criteria remain important even though developers no longer need to oversee every intermediate step.
Version 19.4 extends GitLab's ongoing evolution of its Duo platform throughout 2026, following version 19.2's addition of agents for dependency remediation and security reviews, and earlier releases that introduced flat-rate automated code reviews, free-tier access to the Duo Agent Platform, and monthly credit limits. The /goal command and expanded MCP tools remain in public beta, while the new credit usage visibility is generally available, requiring GitLab 19.3 or later with Duo CLI 9.17.0 or later for Premium and Ultimate customers on GitLab.com, GitLab Self-Managed, and GitLab Dedicated. The shift to longer-running development tasks handled by agents represents a broader industry move to let coding agents act through MCP without granting unrestricted access, part of a pattern also seen in Google's Agent Mode for Android Studio and GitHub Copilot's MCP support with enterprise policies. Organizations now face a balancing act between enabling autonomous workflows that scale development velocity and maintaining visibility into the computational costs those workflows generate, particularly as automation triggers multiply across their pipelines.