
Agent Platform Pipelines, part of Gemini Enterprise Agent Platform, is a managed runner for machine learning workflows that executes pipelines authored with either the Kubeflow Pipelines SDK or the TensorFlow Extended (TFX) SDK. Understanding how these tools divide responsibility is helpful for the Professional Machine Learning Engineer exam, because architectural scenarios often separate pipeline authoring from orchestration and underlying compute infrastructure. Once those three layers are clear, the relationship between Kubeflow Pipelines, TFX, and Agent Platform Pipelines becomes straightforward.
Before Agent Platform Pipelines, teams commonly ran machine learning workflows with Kubeflow Pipelines and the Kubeflow Pipelines Software Development Kit (SDK). That architecture split pipeline work across three layers: authoring, orchestration, and infrastructure.
At the authoring layer, developers write Python code using the Kubeflow Pipelines SDK to define each step in the workflow, such as preprocessing, training, and evaluation, along with how those steps connect. Once defined, the pipeline compiles into a specification that Kubeflow Pipelines can execute. At the orchestration layer, the Kubeflow Pipelines runner reads that specification and executes each step in the proper order using containerized workloads.
Running the Kubeflow Pipelines runner yourself required hosting it on Kubernetes, which is where the name Kubeflow comes from. At the infrastructure layer, teams were responsible for standing up and maintaining clusters, networking, and compute resources, as well as handling scaling and availability. Although Kubeflow Pipelines was a major step forward for structured ML workflows, deploying and maintaining it demanded substantial Kubernetes knowledge just to run machine learning pipelines, which created an operational barrier for many teams.
TensorFlow Extended, or TFX, provides another approach to building production machine learning workflows, particularly for TensorFlow models. Like the Kubeflow Pipelines SDK, TFX operates at the Python-based authoring layer where engineers write pipeline steps. The difference is that TFX is more opinionated and includes a library of standardized, built-in components for common tasks, including ingesting data, transforming it, training a model, and evaluating it.
One distinction worth keeping sharp is that Kubeflow includes both an SDK and a runner, whereas TFX is only the software authoring layer. Referring to TFX or TensorFlow Extended means the TFX SDK. It defines the pipeline steps and logic, not how those steps are orchestrated and executed, which makes the TFX SDK the direct counterpart to the Kubeflow Pipelines SDK.
Because TFX does not include its own execution engine, it requires a separate runner to orchestrate the workflow. Kubeflow Pipelines is the most common runner used with TFX, though TFX also supports other runners such as Apache Beam and Apache Airflow. Regardless of the runner chosen, self-hosted TFX setups still require infrastructure to back that runner, whether on Kubernetes, a virtual machine, or a local machine. Teams therefore still had to configure and maintain both the orchestration system and the environment running it.
Agent Platform Pipelines fits into this stack as a fully managed runner that replaces both the orchestration layer and the infrastructure layer. Nothing changes at the authoring layer. Teams still write pipeline code in Python using either the Kubeflow Pipelines SDK or the TensorFlow Extended SDK, and both SDKs compile down to a standard pipeline specification.
Instead of deploying a self-hosted runner such as Kubeflow Pipelines and managing the Kubernetes clusters beneath it, you pass the compiled pipeline specification to Agent Platform Pipelines. Because orchestration and infrastructure are abstracted away and managed by Gemini Enterprise Agent Platform, there is no Kubernetes cluster setup, no manual scaling, and no separate orchestration engine to maintain. Teams can focus on writing pipeline logic while Agent Platform Pipelines handles execution.
When a Professional Machine Learning Engineer scenario asks how to build and run an ML workflow on Gemini Enterprise Agent Platform, the authoring choice and the execution choice can be evaluated separately:
Keeping the three layers distinct clarifies how these tools relate. The Kubeflow Pipelines SDK and the TFX SDK both live at the authoring layer and compile into a standard pipeline specification, while Agent Platform Pipelines serves as the managed runner that takes over orchestration and infrastructure once the pipeline is defined.
Our Professional Machine Learning Engineer course covers Agent Platform Pipelines with Kubeflow Pipelines and TensorFlow Extended alongside model training and evaluation workflows on Gemini Enterprise Agent Platform, with practice questions that drill these distinctions.