Skip to content

Saving & Rating Workflows

Saving a task to the workflow library does not mean creating a separate export file, but rather saving task evaluation and workflow content together to promote it to a reusable example.

Information Saved

Evaluation Metadata

  • rating
  • useCase
  • workflowTags
  • outcomeNote
  • approachNote
  • reuseWhen
  • watchouts
  • evaluatedAt

Workflow Content

The server storage record also includes:

  • workflowSnapshot
  • workflowContext
  • searchText

searchText is not a separate form field but is saved based on the search string from the snapshot.

Save Paths

Web

Saved through TaskEvaluatePanel in the Evaluate tab of EventInspector.

The web path is not just a simple note input screen.

  1. Receives evaluation input values.
  2. Auto-generates snapshot/context from timeline.
  3. User can preview or edit fields to review and modify generated values.
  4. Can revert to auto-generated values with Regenerate.
  5. On save, sends evaluation + workflowSnapshot + workflowContext to POST /api/tasks/:id/evaluate.

MCP / Agent

Manual saving is also possible via the monitor_evaluate_task tool. Values directly passed in this path are primarily evaluation metadata.

  • rating
  • useCase
  • workflowTags
  • outcomeNote
  • approachNote
  • reuseWhen
  • watchouts

If snapshot/context override is not sent, the server auto-generates and saves based on the current task timeline.

good and skip

  • good: Work worth revisiting next time
  • skip: Kept in library but not prioritized in recommended examples

Sorting generally has good first, with latest evaluatedAt coming first within that.

Where It Is Used After Saving

  • GET /api/workflows list panel
  • GET /api/workflows/similar similar example search
  • GET /api/workflows/:id/content snapshot/context detail view
  • Agent's workflow library search rules

Criteria for Good Evaluation Notes

  • useCase should be short and categorizable.
  • outcomeNote should immediately show what was resolved.
  • approachNote should explain why this approach worked.
  • reuseWhen should indicate when to revisit next time.
  • watchouts should be specific to avoid repeating the same mistakes.

Local-first documentation for Agent Tracer.