How to Orchestrate Gemini Enterprise Agent Platform Pipelines with UAC

A machine learning (ML) pipeline can be fully automated and still leave you with a manual process. 

That’s because the pipeline is rarely the whole process. Data has to arrive first. Another job may need to prepare it. Applications may be waiting for the result. And when something fails, somebody needs enough context to understand what happened and what didn't run as a result.

Gemini Enterprise Agent Platform Pipelines (formerly Google Vertex AI Pipelines) handles the ML pipeline. Stonebranch Universal Automation Center (UAC) can orchestrate that pipeline as part of the larger enterprise workflow. 

Key Takeaways

• UAC can run, stop, and monitor Gemini Enterprise Agent Platform Pipelines (GEAPP) within larger enterprise workflows.
• Runtime parameters let teams reuse ML pipelines while supplying execution-specific context from UAC.
• Pipeline caching can help avoid unnecessary repeat work when eligible outputs can be reused.
• UAC brings pipeline execution into the operational context of the larger workflow, including upstream and downstream dependencies.
• The bigger opportunity is moving from an isolated ML pipeline to an orchestrated enterprise process. 

The Difference Between ML Pipelines and Enterprise Orchestration 

Agent Platform Pipelines is designed to automate and execute multi-step ML workflows. It understands the components of the pipeline and the relationships between them. 

UAC doesn’t replace that. Instead, think one level above the ML flow. There, the Agent Platform Pipeline is only one link in an enterprise-level chain. For example, a production retraining process might look like this:

Data arrives → preparation completes → validation passes → Agent Platform Pipeline runs → results become available → downstream processing begins

This is where the distinction between ML pipelines and enterprise orchestration becomes important. Agent Platform Pipelines manages the steps and dependencies within the ML pipeline. UAC orchestrates the larger process — determining when the ML pipeline should run, what must happen first, and what happens after it finishes.

Keep ML Logic in the ML Platform

Enterprise orchestration doesn’t mean rebuilding the ML pipeline in UAC. Agent Platform Pipelines continues to manage its ML components and dependencies, while UAC coordinates how that pipeline interacts with the rest of the enterprise. 

This lets the ML teams keep working with their preferred tools while operations teams gain centralized control and visibility across the broader process.

As ML moves deeper into production, that broader view becomes increasingly important. Once other applications and processes depend on an ML pipeline, it’s no longer an isolated ML job. It’s part of an enterprise workflow.

Run Gemini Enterprise Agent Platform Pipelines from Stonebranch UAC

Stonebranch Universal Automation Center offers an integration that can launch Gemini Enterprise Agent Platform Pipelines directly from a UAC workflow.

When configuring the task, users can:

• Run or stop a pipeline
• Select the region
• Choose the pipeline template
• Provide optional parameters using scripts or JSON
• Enable caching
• Specify the display name for the pipeline execution

The video below shows the UAC Agent Platform Pipelines task in action:

UAC can also monitor pipeline execution and stop the pipeline when needed. Individual pipeline steps appear within the task instance, and execution details are available in the log after the task finishes. 

Download the UAC Agent Platform Pipeline Task

Four Ways to Orchestrate ML Pipelines in Production

So, what does this look like in practice? Here are four orchestration patterns that address common challenges in production ML.

1. Replace “It Should Be Ready by Then” with a Real Dependency

One surprisingly persistent pattern in enterprise automation is using a schedule to approximate a dependency. Suppose a retraining pipeline needs data produced by an upstream process. The data normally finishes at 1:30am, so the ML pipeline is scheduled for 2am — just to be on the safe side. 

That works… Until it doesn’t. 

A longer-than-usual data load can leave the pipeline starting against incomplete inputs. Adding more buffer reduces the risk but also increases idle time. 

Enterprise orchestration provides a better model: Don’t start because it’s 2am. Start when the required data becomes available. 

With UAC, the successful completion of the upstream process can become a dependency for the Agent Platform Pipelines task. In addition to traditional time-based scheduling, UAC triggers automation workflows based on system events – powering real-time automated processes. 

The workflow becomes:

Data ready → validation complete → ML retraining pipeline starts

The ML pipeline runs when its required inputs are actually available. For production MLOps, that’s a meaningful shift. 

2. Reuse Pipelines Without Rebuilding Workflows

The UAC Agent Platform Pipelines task can supply optional parameters as scripts or JSON. 

That establishes a useful architectural pattern: let the ML team define what the pipeline does, while the enterprise workflow supplies the context for a specific run.

Depending on how the pipeline is designed, that context might represent different datasets, processing dates, environments, or other inputs.

The result is a cleaner operating model. ML teams maintain the pipeline logic, while automation teams control when it runs and how it fits into the larger enterprise workflow.

3. Use Caching Where It Makes Sense

The Stonebranch task for Agent Platform Pipelines also supports pipeline caching.

When caching is enabled, eligible outputs from previous pipeline steps can be reused rather than repeating the same work. That can be valuable for ML flows with computationally expensive stages. But caching shouldn’t be simply treated as a switch to turn on everywhere. 

The obvious benefit is avoiding unnecessary work. But the more useful question is whether a particular step actually needs to run again. If the relevant inputs haven’t changed, reusing an eligible output may save time and compute resources. 

That makes caching part of the workflow conversation. A useful separation is: UAC determines whether the overall process should execute. Agent Platform Pipelines determines which eligible work within the pipeline can be reused. 

4. See the Pipeline in the Context of the Whole Workflow

Google’s platform already provides visibility into ML pipeline execution. So why monitor it from UAC too? Context.

Imagine this sequence:

1. Data ingestion succeeds.
2. Preparation succeeds.
3. The ML pipeline starts.
4. A pipeline step fails.
5. A downstream application update never runs. 

An ML engineer needs to know which pipeline component failed. The operations team needs to know where the end-to-end process stopped, and what was affected downstream.

Stonebranch exposes pipeline steps in the UAC task instance and captures execution details in the task log. That puts the failure in context, alongside the processes that ran before it and those waiting downstream.

That’s particularly useful when the team operating the overall workflow isn’t the team that built the ML pipeline. 

Bring ML Pipelines into Your Enterprise Workflows

Gemini Enterprise Agent Platform Pipelines provides the environment to execute the ML workflow. Stonebranch UAC connects that workflow to the data, applications, infrastructure, and processes around it. 

The result is a simple division of responsibilities: keep the ML logic where it belongs and use enterprise orchestration to connect it with the rest of your IT environment. 

Want to check out how this works for yourself? Download the AI Agent Pipeline integration now.