GitLab launches agent orchestration: the /goal command drives a task to green CI

GitLab is trying to eliminate the waiting between stages of development. At its Transcend conference, the company announced three new agent orchestration features in Duo Agent Platform: goal-driven flows, the GitLab for Slack integration and Custom Flows. Goal-driven flows require GitLab 19.3 and Duo CLI 9.17.0 or newer. GitLab's argument is simple: AI has sped up writing code, but not delivering it. An agent produces a fix in minutes, and then the fix waits weeks for a reviewer's sign-off, for someone to work through pipeline alerts, for scan results to be triaged, and for yet another manual prompt.

A goal instead of steps: how /goal works

Goal-driven flows give an agent a finish line rather than the next action. The developer describes the task and the criterion for "done," and the flow keeps working until it reaches the result. Review, tests, security checks and approvals run as one connected chain, with no stops between steps.

Launching one looks like this: `/goal Fix the checkout timeout. Continue until tests pass and CI is green.` The command works in Duo CLI, in headless mode via `duo run --goal`, and in Agentic Chat. The headless form matters for automation: it can be called from scripts and pipelines without an interactive session.

Quality control gets its own mechanism. At every step, a separate model checks the work against the goal, so the agent does not grade itself. A run can be stopped or redirected at any moment.

Slack and Custom Flows: start where people ask

GitLab for Slack addresses the entry-point problem. Most tasks originate in a discussion, not in a tracker. You mention @GitLab in a channel or thread and write something like "fix the failing pipeline": a flow starts, and Duo streams progress straight into the thread. The same bot answers questions about the codebase and turns conversations into issues. The flow runs under the same identity and policies as any other flow in GitLab, so access control stays with the platform, not the chat.

Custom Flows cover the third scenario: a team defines its own multi-step sequence of agents and ties it to GitLab events such as a mention, an assignment or a pipeline. Flows execute in CI/CD. GitLab's example is a self-healing pipeline that analyzes a failed test, generates a fix, commits it and notifies the team. Sensitive steps can get human checkpoints.

All of this runs under a single governance model: identity, policies and the evidence chain are shared across all flows. For companies that face audits, this is arguably the main argument for a platform approach over a scatter of external agents. The set can be extended through the AI Catalog, which connects external agents and tools via the MCP client and MCP server. Built-in agents and flows already cover stages from planning to analysis in the GitLab interface, the IDE and Duo CLI.

Availability and first steps

Launch terms:

• goal-driven flows: GitLab.com, Self-Managed and Dedicated, Premium and Ultimate tiers;
• minimum versions: GitLab 19.3 and Duo CLI 9.17.0;
• Custom Flows: already generally available (GA);
• MCP server and GitLab for Slack: GA expected later this month;
• billing: GitLab Credits, which are included in a Premium or Ultimate subscription.

It makes sense to start small. First check your GitLab and Duo CLI versions, enable Duo Agent Platform, and run `/goal` on a narrow task with an unambiguous completion criterion, such as a failing test with a reproducible error. The wording of the goal decides almost everything: "green CI" can be verified by a machine, "improve the code" cannot. Then move a repetitive manual process into a Custom Flow, keeping human checkpoints on merges and on anything involving secrets. It makes sense to plan the Slack connection for after GA, when the integration leaves early access.

The main question remains open: how reliably the checking model can tell a truly completed goal from a pipeline that is merely green on paper. That is worth measuring on your own repositories before flows are allowed to move code past review.