Workflow
The workflow is fixed and agent-centered. The human provides the brief, reviews the recommendation for each step, and can revise the active policy through Action mode at any point.
Canonical workflow
The canonical workflow lives in:
src/agentic_automl/assets/automl_workflow.md
The steps are:
Intake
Preprocessing
Data splitting
Model selection
Metric selection
Training configuration
Validation and baseline
Hyperparameter optimization
Final validation
Per-step assets
Each workflow step ships with three companion files under
src/agentic_automl/assets/skills/<step>/:
SKILLS.mddescribes how the step reasons and operates.KNOWLEDGE.mddeclares the currently supported executable actions.LIMITS.mdstores unsupported requests and seed backlog items.
Step intent
Intake captures the minimum project brief.
Preprocessing owns data cleaning, feature pruning, feature-role changes, and executable preprocessing overrides.
Data splitting owns the final holdout strategy only.
Model selection chooses one specific starting model and its initial parameters.
Metric selection chooses one winner metric that also governs baseline comparison.
Training configuration controls the executable training parameters relevant to the selected model.
Validation and baseline compare the current model against the strongest simple no-feature baseline.
Hyperparameter optimization optionally runs a focused competition on the selected model.
Final validation summarizes the tuned-versus-untuned outcome and prepares the final notebook story.