## Release notes Creating a valid custom flow on the GitLab Duo Agent Platform now takes a single conversation with a purpose-built foundational agent, instead of hand-authoring YAML against the Flow Registry schema. Previously, flow authors had to know component types, input syntax, context variables, routing patterns, and known failure modes up front. Small mistakes produced YAML that failed silently or with opaque errors, turning flow creation into a debugging exercise. The new Flow Creator foundational agent reads the current Flow Registry v1 documentation on every response, applies hard rules that encode real-world failure patterns, and runs a pre-output checklist before generating YAML. The result is a complete, runnable flow definition you can register and run immediately from the AI Catalog -- even on your first attempt. https://docs.gitlab.com/ee/user/duo_agent_platform/flows/custom/ --- ### Overview Enable GitLab customers to create valid Custom Flows using a purpose-built Foundational agent. ### Problem Building custom flows on the Duo Agent Platform requires precise knowledge of component types, input syntax, context variables, routing patterns, and known failure modes. Without guardrails, flow authors frequently produce broken YAML that fails silently or with opaque errors — causing lost time in debugging rather than building. ### Solution A foundational Flow Creator agent, powered by an optimized system prompt, that: * Reads the current Flow Registry v1 documentation before every response to stay current as the framework evolves * Applies a set of hard rules encoding real-world failure patterns (missing `project_id`, broken HITL gates, wrong `sends_response_to` targets, missing stopping instructions) * Runs a pre-output checklist before generating any YAML to catch common mistakes before they reach a session * Produces complete, runnable YAML — never partial snippets — that a flow author can register and run immediately ### Success criteria * A flow author with no prior Flow Registry experience can produce a working YAML on the first attempt using the agent * The agent correctly identifies and names rule violations when debugging a broken flow * The agent stays current with framework changes by reading live documentation rather than relying on training knowledge