Kedro: version 1.2.0 brings the @experimental decorator and a LangGraph agentic starter for GenAI pipelines

The Kedro project, part of the Linux Foundation AI ecosystem, released a dual release on May 22, 2026: Kedro 1.2.0 (core framework) and Kedro-Viz 12.3.0 (visualization tool). The version brings three significant new capabilities: the @experimental decorator for APIs under development, an official LangGraph integration starter project, and an expanded node preview system in Kedro-Viz with support for Mermaid diagrams, text, and images.

What is the @experimental decorator and how is it used?

@experimental is a new Python decorator that marks a function or class in the Kedro API as “experimental.” The marked API is publicly available, but the Kedro core team explicitly reserves the right to change its signature, behavior, or remove it without a breaking change process.

Users who call an experimental API receive a runtime warning through the Python warnings module (DeprecationWarning subclass ExperimentalWarning):

ExperimentalWarning: kedro.pipeline.langgraph_node is experimental and may change.

The warning can be filtered through the standard warnings.filterwarnings() if a team uses an experimental feature in production and does not want log noise. The decorator’s goal is twofold: encouraging the Kedro core team to iterate quickly on new ideas without fear of breaking changes, and clearly informing users that a particular API is not yet stable.

Which are the first @experimental APIs in 1.2.0?

The first two APIs marked with @experimental are LLM context nodes and agentic workflow primitives. An LLM context node is a specialized Kedro node that does not perform classical data transformation but calls an LLM (Claude, GPT, Gemini) with a structured prompt template from Kedro config. The output is returned as a Python dict that enters the next node — enabling hybrid pipelines where some phases do classical ML processing and others do LLM reasoning.

Agentic workflow primitives allow an entire Kedro pipeline to be treated as an agent that autonomously decides which node to execute based on runtime state. This is explicitly experimental because it changes Kedro’s semantic foundation (pipeline as a DAG with fixed ordering) into something more flexible but less predictable.

What is the support-agent-langgraph starter?

Kedro starters are pre-configured project templates cloned with kedro new --starter=support-agent-langgraph. The new LangGraph starter shows a real-world enterprise scenario: a customer support agent that reads Slack notifications, extracts context from Confluence and CRM, generates a draft response through an LLM, and routes to a human agent if confidence falls below a threshold.

The technology stack is instructive:

• Kedro pipeline for data ingestion (Slack RTM, Confluence API, CRM connectors)
• LangGraph for orchestrating agent logic (decision tree, retry handling, escalation)
• Langfuse and Opik (both open-source) for prompt versioning, evaluation, and observability

The starter comes with example data, a complete Docker Compose setup, and a Kedro test suite. The goal is not to be production-ready out of the box but to show which components need to be connected for an enterprise GenAI workflow with Kedro as the backbone.

What is new in Kedro-Viz 12.3.0?

Kedro-Viz is the web-based visualization tool for Kedro pipelines. Version 12.3.0 significantly expands the “node preview” system — functionality that shows a summary of each node in the browser (input/output, metadata, last run status).

New features:

• Mermaid diagram preview: if a node contains Mermaid markdown in its documentation, Kedro-Viz renders it inline as an SVG diagram. Useful for documenting internal ML algorithms or business logic that is more easily understood visually.
• Text and image preview: a node can return text (sample output) or an image (e.g., a matplotlib plot) displayed in the side panel without manually saving files.
• Custom preview extensibility: developers can register custom previewers for their own data types through the Kedro-Viz plugin API.

Multi-pipeline execution: the addition is not new in 12.3.0 but has been expanded — multiple independent Kedro pipelines can now be run in parallel from the Kedro-Viz UI with real-time progress tracking. Useful for enterprise tenants running 10–50 ML pipelines simultaneously.

What does this mean for the Kedro user base?

Kedro has reached the milestone of 30,000 GitHub stars and has approximately 200 active contributors through LF AI governance. The 1.2.0 release clearly signals a pivot toward GenAI workflow support — Kedro is no longer exclusively a classical ML framework but an MLOps platform that covers both traditional ML and GenAI agentic pipelines in the same foundation. This is strategically significant as it opens Kedro adoption to enterprise tenants who previously had to choose between Airflow (classical ML) and LangChain (GenAI) — Kedro offers a single source.