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:

  1. Intake

  2. Preprocessing

  3. Data splitting

  4. Model selection

  5. Metric selection

  6. Training configuration

  7. Validation and baseline

  8. Hyperparameter optimization

  9. Final validation

Per-step assets

Each workflow step ships with three companion files under src/agentic_automl/assets/skills/<step>/:

  • SKILLS.md describes how the step reasons and operates.

  • KNOWLEDGE.md declares the currently supported executable actions.

  • LIMITS.md stores 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.