Latest

By news editor

Google Cloud has made its Managed Airflow MCP server generally available, which lets any AI agent that supports the Model Context Protocol manage a customer's data pipelines on the company's Managed Service for Apache Airflow. The Google Cloud post that carries the news, by senior product manager Piotr Wieczorek and software engineer Khadija Al Ahyane, also walks through a console agent that diagnoses failed pipeline runs and a kit that lets coding agents write and deploy pipelines.

Apache Airflow is the open-source tool many data teams use to schedule and run pipelines, which it calls DAGs. Google's managed version was formerly named Cloud Composer, the post notes. The Model Context Protocol, or MCP, is an open standard for connecting AI agents to outside tools and data.

Through the MCP server, an agent can list and trigger DAGs, check the status of runs and inspect failed tasks, according to the post. Access is governed by Google Cloud's identity and access management controls, known as IAM. Tools that only read data are flagged, so an administrator can limit an agent to those.

A second piece, Managed Airflow Agent, sits in the Google Cloud console. The post describes three jobs for it: investigating a failed run, recommending changes to an Airflow environment, and checking whether existing DAGs are compatible with a newer Airflow version before an upgrade. An investigation starts from a button in a pipeline's run history. In the failure case the post walks through, the agent correlates worker memory metrics with task retries and proposes fixes.

That agent does not act on its own findings. The post says it is read-only, inspects production telemetry with the user's own permissions and does not change anything. A person makes the change.

Coding agents get a separate route. Data Agent Kit adds Google data skills and tools to Antigravity, Claude Code, Gemini CLI and Codex CLI, and also comes as a VS Code extension, the post says. For Airflow, it can write a DAG that matches the Airflow version of the target environment, deploy it, and migrate existing DAGs to Airflow 3. Claude Code is made by Anthropic and Codex CLI by OpenAI, so the kit is not limited to Google's own agents. Google has also published its Airflow agent skills as open source.

Beyond its own products, the post points to the Apache Airflow Common AI provider, linking to the Airflow project's documentation. The provider lets model-driven steps run beside conventional ones in an Airflow 3 pipeline. The examples given are branching decided by a large language model, retries that separate temporary errors from bad input, and detection of schema drift, meaning unexpected changes in the structure of incoming data.

What the post does not contain is evidence from a customer. Its worked example follows Cymbal Shops, which Google states is a fictional company, and a closing note says the agent outputs and the traffic and memory numbers in that story are shortened and illustrative. No customer is named. No price is given for the MCP server, the console agent or the kit.

Release status is also incomplete. The MCP server is the only item the post calls generally available; it does not say whether the console agent and Data Agent Kit are generally available or in preview. Nor does it say which model powers the console agent, or how often the agent's diagnoses are correct.

The next thing to watch is whether Google attaches a price or a usage charge to the agents, and whether it publishes results from a named customer running them against real pipeline failures. Release notes for Managed Service for Apache Airflow would show when the console agent and the kit reach general availability.

Leave a Reply