What does GitLab Duo Agent Platform actually do for developers?

GitLab Duo Agent Platform embeds multiple intelligent assistants across the software development lifecycle. Instead of a linear workflow you collaborate asynchronously with specialized agents.

Availability depends on tier and version

The platform runs on GitLab.com, on GitLab Self-Managed, and on GitLab Dedicated. It requires Premium or Ultimate tier. The underlying model is Anthropic Claude Sonnet 4. Self-hosted models are also supported.

History moves fast. Beta arrived in GitLab 18.2. Self-managed instances got an experimental version in 18.4 behind a feature flag named self_hosted_agent_platform. That flag was disabled by default. It became enabled in 18.7. General availability landed in 18.8. GitLab Credits support started the same release. The flag was removed in 18.9.

You cannot use the Agent Platform with the GitLab Duo Enterprise add-on on 18.9 or earlier. Upgrade to 18.10 or later for that combination.

Prerequisites are specific

Duo must be turned on. If you lack Duo Pro or Enterprise you need Duo Core enabled for the top-level group or instance. On 18.8 and later the Agent Platform itself must be turned on. On 18.7 and earlier you need beta and experimental features enabled.

Self-managed instances require instance configuration. GitLab Duo Self-Hosted demands a self-hosted AI Gateway with the Agent Platform service. Local use needs an editor extension authenticated with GitLab. Your project must live in a group namespace. A Developer, Maintainer, or Owner role is required.

Generally available features consume credits

Most features burn GitLab Credits when used. Free tier access requires purchasing credits. The table below shows what ships as generally available.

Agentic Chat answers complex questions and autonomously creates and edits files. Code Suggestions delivers AI-powered suggestions as you write. Custom agents let teams build agents for unique requirements. External agents connect third-party integrations securely — Premium and Ultimate only.

Planner Agent plans, prioritizes, and tracks work. Data Analyst Agent analyzes metrics and project data. Developer Flow converts issues into merge requests. Code Review Flow automates review tasks and enforces standards. Convert to GitLab CI/CD Flow transforms legacy pipelines. Fix CI/CD Pipeline Flow diagnoses and automatically fixes failing pipelines.

Software Development Flow creates a full multi-step plan in your IDE before executing. MCP clients access GitLab resources from any MCP-compatible client or IDE extension. Custom flows combine multiple agents. Interactions in merge requests let you ask questions and discuss feedback in comment threads.

Resolve merge conflicts autonomously analyzes conflicts, edits files, and pushes a resolution commit — Premium and Ultimate. Recommend Reviewers Flow suggests the best reviewers — Premium and Ultimate. Flow Creator Agent builds custom flows for the AI Catalog — Premium and Ultimate. Resolve review discussions analyzes discussions, pushes changes, and resolves threads — Premium and Ultimate.

SAST False Positive Detection Flow filters false positives in SAST scans — Ultimate only. SAST Vulnerability Resolution Flow generates fixes for SAST vulnerabilities — Ultimate only. Secret False Positive Detection Flow filters false positives in secret detection — Ultimate only. Permissions Assistant selects correct permissions for fine-grained personal access tokens — Ultimate only. Security Analyst Agent automates triage, vulnerability analysis, and fix generation — Ultimate only. MCP server connects AI tools securely to your GitLab instance.

Beta and experimental features split on credit consumption

Some beta features consume credits. GitLab for Slack app lets you interact with Duo from Slack — Premium and Ultimate. Security Review Flow detects business logic vulnerabilities in merge requests — Ultimate only. Workspace subagents define specialized subagents in your repository for Developer Flow delegation.

Other beta features do not consume credits today. Agent tool governance configures tool-level approval policies for sensitive actions. AI audit event report provides a unified record of agent activity — Premium and Ultimate. AI Governance Dashboard monitors sessions, audit logs, and developer exposure across a group — Ultimate only. CI Expert Agent creates, debugs, and optimizes pipelines. External MCP servers connect custom agents to external data sources — Premium and Ultimate. Agentic Breaking Change Resolution Flow analyzes pipeline failures on dependency bump merge requests and creates code fixes — Ultimate only. Support Assistant diagnoses and resolves GitLab product problems — Premium and Ultimate.

MCP clients do not consume credits directly. However any Agent Platform usage such as model requests made through an MCP client might consume credits.

What changes for the person writing code

Developers delegate routine work. Refactoring, security scans, research — specialized agents handle them asynchronously. The platform does not force a single linear path. You can invoke a Planner Agent to structure work, a Developer Flow to turn an issue into a merge request, a Code Review Flow to enforce standards, and a Fix CI/CD Pipeline Flow to repair broken pipelines. All of this happens inside the same GitLab environment you already use.

Credits are the new currency. Every generally available feature draws from your monthly commitment pool. Free tier users must buy credits upfront. Beta features that currently cost nothing may start consuming credits the moment they graduate to general availability. The documentation states this transition can happen at any time.

Administrators on self-managed instances face extra steps. The feature flag self_hosted_agent_platform existed only between 18.4 and 18.8. It is gone now. Current self-managed installs need the AI Gateway with the Agent Platform service configured. Cloud-connected instances skip that but still require Duo enabled and the platform toggle on.

Enterprise add-on users on older versions are blocked. If you run GitLab Duo Enterprise on 18.9 or earlier the Agent Platform simply will not work. The fix is a version upgrade to 18.10 or later.

The document does not say how many credits each agent action costs. It does not publish latency numbers or accuracy benchmarks. It does not compare the platform against GitHub Copilot or other assistants. It lists capabilities and prerequisites — nothing more.