Pipelines

Pipelines chain Advanced Training work inside one project: train an experiment, optionally deploy a checkpoint to Agents, then run the next stage by hand or when the previous stage succeeds.

Classic RL projects do not show a Pipelines tab. The choice is fixed at project creation via Advanced Training. See Classic vs Advanced Training.

Opening the tab

On the project page, Pipelines appears only for Advanced Training projects. Tab order: Experiments, Pipelines, Results.

Pipeline list

New Pipeline opens a modal (name required, description optional). New pipelines start with manual stage triggers unless you change stage settings later.

The table columns:

Column	Notes
Name	Inline rename via row menu Edit name
Last modified	Latest change to the pipeline or any stage
Actions	Duplicate, Delete (with confirmation)

Expand a row to edit the description and open the pipeline builder graph.

Empty state: No pipelines found with a prompt to create one.

Stages

Each pipeline is an ordered list of stages. Two types appear in the stage dialog:

Type	Purpose
Experiment	Run or resume an Advanced Training experiment
Deployment	Deploy a checkpoint from the previous experiment stage to Agents

Deployment cannot be the first stage.

Add stages from the graph. You can insert at a position, reorder, edit settings, delete, and Trigger stage on manual stages. The builder refreshes about every five seconds so status badges update without a full page reload.

Trigger

Setting	Behavior
Manual	Runs only when you trigger the stage (required for the first stage)
Auto	Starts when the previous stage completes successfully

Copy on the form: First stage can only be triggered manually. Auto stages explain that they start when the prior stage succeeds.

Checkpoints between stages

Experiment stages after the first:

• Best checkpoint or Final checkpoint from the previous stage

• Optional Auto-add checkpoint to Agents when complete

• LatentPPO: helper text that the previous checkpoint is used as the decoder

First Experiment stage:

• Start from scratch, or

• Use saved Agent (pick from Saved Agent dropdown)

Deployment stages:

• Checkpoint from Previous Stage: Best checkpoint or Final checkpoint only

Experiments on stages

Stages bind to an experiment in the same project. The experiment picker filters by compatibility with the previous stage’s algorithm. If the experiment already ran or is used elsewhere in the pipeline, the UI offers duplication before you attach it.

From an experiment stage node you can open Results for that experiment. Stage bubbles show type, status, and Stale when a prior stage was re-run.

Limits

• Classic RL projects have no tab and cannot use pipelines.

• Stages stay inside one project; there is no cross-project chain.

• Pipelines do not replace the experiment wizard; configure each experiment there, then wire stages to those experiments.

To compare metrics without a pipeline, use the Results tab.