GitLab 19.4 Adds Agentic Automation With Cost Controls

GitLab 19.4 Adds Agentic Automation With Cost Controls

GitLab has released version 19.4 of its DevSecOps platform, introducing tools that let organizations scale AI agent automation while maintaining cost control and governance. The update addresses a key barrier to enterprise adoption: giving platform owners the visibility and controls needed to extend agentic workflows beyond individual developers to entire engineering teams.

According to GitLab, the release was first detailed by Business Wire on September 18, 2026.

Goal-Based Automation in the Terminal

The centerpiece of the release is a new /goal slash command in GitLab Duo CLI, now in public beta. Unlike existing AI agents that handle discrete tasks, the command lets developers delegate an entire objective. The agent implements the work locally under existing project guardrails, while a separate model verifies each step against the stated goal. The flow continues until the objective is met, the iteration limit is reached, or the developer stops it.

Developers can revise goals mid-run and restart, with all work governed by the organization's existing permission model. The verification layer provides an independent check on agent judgment, addressing concerns about handing off open-ended work without supervision.

Open-Weight Models Cut Costs Up to 4x

GitLab Duo Agent Platform now hosts three open-weight models—Kimi K3, MiniMax M3, and GLM 5.3—alongside existing frontier models. The new options deliver up to four times more API calls per GitLab Credit compared to many frontier models, at comparable performance levels.

Group owners can set default models for each feature and curate which models teams can access, with settings cascading to child groups and projects. GitLab vets each model against internal performance standards and evaluates hosting vendors through its third-party risk management process.

MCP Tools Extend Automation Surface

New tools in GitLab's Model Context Protocol server, now in public beta, let agents working in external MCP clients automate complete workflows across GitLab. Agents can trigger CI/CD pipelines, read job traces, manage merge requests from opening through merge, search and update work items, and triage vulnerabilities.

Administrators govern these tools through the same GitLab Duo Agent Platform rules already configured in group and project settings. Read-only tools default to "Always Allow" for routine lookups, while write and delete operations default to "Always Ask," giving reviewers a checkpoint before changes.

Usage Visibility for Platform Owners

GitLab 19.4 makes credit consumption visible at the user level, now generally available. Platform owners can set per-user caps and export usage data down to individual billable events. Exports arrive by email with secure download links for both GitLab Flex and non-Flex subscriptions. Developers can also view their own consumption for the first time, enabling teams to self-manage their automation pace.

Additional Capabilities

The release includes GitLab Duo Agent Platform integration with Slack as an experiment for Premium and Ultimate customers. Teams can mention @GitLab in threads to search projects, open issues, or get answers using thread and channel history as context. The agentic flow runs on a CI/CD runner and posts results back to the thread.

Model selection for the Developer Flow in Duo Agent Platform is now generally available, and a redesigned session details panel surfaces status, timestamps, and triggering users while separating what started an agent session from what it produced.

Why it matters

As AI agents move from developer productivity tools to infrastructure-level automation, enterprises need governance frameworks that scale without creating separate permission models or audit trails. GitLab's approach ties agent capabilities directly to existing code permissions and traces every credit to the user account that spent it, addressing the administrative overhead that has slowed enterprise adoption of agentic workflows. The addition of cost-efficient open-weight models gives organizations a practical lever for matching model expense to task complexity across thousands of automation runs.

"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," said Manav Khurana, chief product and marketing officer at GitLab.

The details were first reported by Business Wire.