DRAFT: docs: Key terms for agentic workflows

docs/agentic/index.md

@@ -0,0 +1,14 @@
+# About Agentic Workflows
+
+_Agentic workflows_ are built from a shared _step_ catalog under `src/nemotron/steps/`.
+Each step declares inputs, outputs, and parameters.
+_Patterns_ document how to combine steps safely.
+Agents and skills, including [**Customize Nemotron**](../customize/index.md), can use the catalog to plan work and generate multi-stage projects.
+
+Use [**Key terms**](key-terms.md) for definitions (*step*, *stage*, *pipeline*, *pattern*, *artifact type*) and to disambiguate recipe stages from catalog steps.
+
+## See also
+
+- [Customize Nemotron](../customize/index.md) — step library in the docs, patterns, models, and contribution guides
+- [Nemotron CLI](../nemotron/cli.md) — recipe commands and global options
+- [Training recipes](../index.md#available-training-recipes) — official Nano3, Super3, Omni3, and Embed flows

docs/agentic/key-terms.md

@@ -0,0 +1,46 @@
+# Key Terms and Concepts
+
+Building agentic workflows using this software requires understanding these terms and concepts.
+
+## A
+
+(artifact-type)=
+Artifact type
+: An artifact type names a class of data or checkpoint passed between _steps_, such as `training_jsonl`, `checkpoint_hf`, `packed_parquet`, and `translated_jsonl`.
+By standardizing the artifacts, you can maintain the interface contract between the steps.
+Types are centralized in `src/nemotron/steps/types.toml` and summarized under [artifact types](../customize/steps/types.md).
+
+## P
+
+(pattern)=
+Pattern
+: A pattern is cross-cutting guidance for composing _steps_ such as when to run eval, how to choose translation backends, packing and tokenizer checks, and so on.
+On disk, patterns are in the `src/nemotron/steps/patterns/` directory and are published under [patterns](../customize/patterns/index.md).
+
+(pipeline)=
+Pipeline
+: A pipeline is the agentic workflow that is implemented as an ordered composition of stages.
+A pipeline defines the multi-step workflow from inputs, such as raw or filtered JSONL, to outputs like checkpoints, eval artifacts, or exports.
+Generated scaffolds describe this in `README`, wire stages from a root **`pipeline.py`**, and may record a canonical graph in **`.generated/pipeline.toml`**.
+
+  This is **not** the same as:
+
+  - **Pipeline parallelism** in Megatron (`pipeline_model_parallel_size`, and related knobs), or
+  - The **NeMo RunSpec** module `nemo_runspec.pipeline`, which schedules jobs for recipe commands.
+
+## S
+
+(stage)=
+Stage
+: A stage is one unit of work in a project.
+On disk, a stage is a directory in the `stages/` directory with a thin `run.py`, configuration files, and an entry for the project CLI.
+Each stage is instantiated from a step ID, such as `sft/megatron_bridge`, plus user-specific settings.
+This meaning of stage is not the same as the training recipies stages such as the Nano3 SFT stage.
+
+(step)=
+Step
+: A step is a reusable building block.
+A step has typed inputs and outputs with a `step.toml` manifest and an optional implementation in `step.py`.
+Steps are grouped by concern, such as convert, curate, prep, sft, translate, eval.
+The canonical list is in the [step library](../customize/steps/index.md).
+Some steps are exposed as `nemotron steps` subcommands of the CLI.

docs/index.md

@@ -171,6 +171,14 @@ nemotron/embed/README.md
 nemotron/artifacts.md
 ```
 
+```{toctree}
+:caption: Agentic Workflows
+:hidden:
+
+About <agentic/index.md>
+agentic/key-terms.md
+```
+
 ```{toctree}
 :caption: Customization
 :hidden: