Example: the content review assistant

Shipped with Studio under Automations, switched off. It shows an agent that chooses a branch, a person approving in the Inbox, an agent working with tools to change a node, and a conversation whose result is written back.

What it does

1. A page in stories is updated.
2. Review the page — the content reviewer agent reads the title and description and picks one of three branches by their labels.
3. It is fine → a task tells the editor.
4. It has an obvious problem → Approve the fix in the editor’s Inbox → Apply the fix, an agent working a tool loop with update-node → a task says what changed.
5. It is unclear → Talk it through, a conversation with the editor that agrees on a title and description → Apply what was agreed writes them → a task says so.

Ignore changes made by automations is on: the fix the agent applies is an update to a page, and without the guard it would start a review of the fix.

The YAML

The complete node, as shipped (content/automations/content-review/.node.yaml). Read it top to bottom with the graph above.

node_type: studio:Automation
archetype: studio:AutomationPage
properties:
  title: Content review assistant
  enabled: false
  trigger:
    kind: node_event
    workspaces: [stories]
    node_types: [studio:Page]
    event_kinds: [Updated]
    ignore_automation_writes: true
  steps:
    - uuid: review
      element_type: studio:StepAgent
      label: Review the page
      agent: { raisin:ref: /agents/content-reviewer, raisin:workspace: functions }
      decide: true
      min_confidence: 0.5
      instruction: |-
        A page was just edited. Review its title and description.

        Title: ${steps.__subject.subject.title}
        Description: ${steps.__subject.subject.description}
        Path: ${steps.__subject.subject_path}

        Choose the branch that fits. A clear, specific title with a useful
        description is fine as it is. An obvious problem you could fix yourself
        — a typo, an empty or vague description, a title that repeats the
        description — needs a fix a person should approve first. If you cannot
        tell what the page is for, talk it through with the editor.
    - uuid: approve_fix
      element_type: studio:StepHumanTask
      label: Approve the fix
      assignee_mode: specific
      assignee: { raisin:ref: /users/internal/admin-at-studio-local, raisin:workspace: raisin:access_control }
      task_type: approval
      prompt: |-
        The content reviewer wants to fix “${steps.__subject.subject.title}”
        (${steps.__subject.subject_path}).

        Its reasoning: ${steps.review.reasoning}

        Approve to let it apply the change, or reject to leave the page as it is.
      options:
        - { label: Approve the fix, value: approve }
        - { label: Leave it, value: reject }
      due_in_hours: 48
    - uuid: apply_fix
      element_type: studio:StepAgent
      label: Apply the fix
      agent: { raisin:ref: /agents/content-reviewer, raisin:workspace: functions }
      tools: true
      max_iterations: 6
      instruction: |-
        The fix you proposed was approved. Apply it now.

        The page is at ${steps.__subject.subject_path} in the workspace
        "${steps.__subject.workspace}". Its current title is
        "${steps.__subject.subject.title}" and its description is
        "${steps.__subject.subject.description}".

        Use update-node to change only what is wrong (workspace
        "${steps.__subject.workspace}", path "${steps.__subject.subject_path}").
        Then say in one paragraph what you changed.
    - uuid: talk
      element_type: studio:StepChat
      label: Talk it through
      agent: { raisin:ref: /agents/content-reviewer, raisin:workspace: functions }
      participant_mode: specific
      participant: { raisin:ref: /users/internal/admin-at-studio-local, raisin:workspace: raisin:access_control }
      system_prompt: |-
        You are talking with the editor about the page
        “${steps.__subject.subject.title}” at ${steps.__subject.subject_path}.
        You could not tell what it is for. Ask what the page should say, agree
        on a title and a one-sentence description, and finish the conversation
        with them as your result.
      max_turns: 12
      allow_ai_end: true
      output_schema:
        type: object
        properties:
          title: { type: string, title: Agreed title }
          description: { type: string, title: Agreed description }
        required: [title]
    - uuid: apply_agreed
      element_type: studio:StepFunction
      label: Apply what was agreed
      function_ref: { raisin:ref: /lib/studio/automations/library/update-node, raisin:workspace: functions }
      arguments:
        workspace: ${steps.__subject.workspace}
        path: ${steps.__subject.subject_path}
        properties:
          title: ${steps.talk.result.title}
          description: ${steps.talk.result.description}
    - uuid: notify_agreed
      element_type: studio:StepFunction
      label: Tell the editor what was agreed
      function_ref: { raisin:ref: /lib/studio/automations/library/create-task, raisin:workspace: functions }
      arguments:
        assignee: /users/internal/admin-at-studio-local
        title: 'Updated after our chat: “${steps.talk.result.title}”'
        task_type: review
    - uuid: notify_ok
      element_type: studio:StepFunction
      label: Tell the editor it is fine
      function_ref: { raisin:ref: /lib/studio/automations/library/create-task, raisin:workspace: functions }
      arguments:
        assignee: /users/internal/admin-at-studio-local
        title: 'Reviewed: “${steps.__subject.subject.title}” looks good'
        description: ${steps.review.reasoning}
        task_type: review
    - uuid: notify_done
      element_type: studio:StepFunction
      label: Tell the editor what changed
      function_ref: { raisin:ref: /lib/studio/automations/library/create-task, raisin:workspace: functions }
      arguments:
        assignee: /users/internal/admin-at-studio-local
        title: 'Reviewed and updated: “${steps.__subject.subject.title}”'
        description: ${steps.apply_fix.response}
        task_type: review
  edges:
    - { id: e_fine, source: review, target: notify_ok, kind: next, label: it is fine }
    - id: e_fix
      source: review
      target: approve_fix
      kind: condition
      label: The page has an obvious problem I can fix — a typo, an empty or vague description — and a person should approve the fix first
    - id: e_talk
      source: review
      target: talk
      kind: condition
      label: I cannot tell what this page is for and need to talk it through with the editor
    - { id: e_approved, source: approve_fix, target: apply_fix, kind: condition, condition: "steps.approve_fix.action == 'approve'", label: approved }
    - { id: e_applied, source: apply_fix, target: notify_done, kind: next }
    - { id: e_talked, source: talk, target: apply_agreed, kind: next }
    - { id: e_agreed, source: apply_agreed, target: notify_agreed, kind: next }
  entries: [review]
  exits: [notify_ok, notify_done, notify_agreed, approve_fix]

Things to notice

• The router has no schema. The three connections out of Review the page carry the sentences the agent chooses between; the plain next one is the fallback under the minimum confidence.
• Three ways to read an agent. steps.review.reasoning from the router, steps.apply_fix.response from the tool loop, steps.talk.result.title from the conversation.
• The approval gates the write. Apply the fix only runs on steps.approve_fix.action == 'approve'; “Leave it” ends the run with the page untouched.
• The agent’s tools are on the agent. The content reviewer lists find-nodes and update-node; the step only says let it work a tool loop.

Try it

Enable it, edit any page’s description in stories, and watch the Inbox. To point it at your own editor, change the assignee and participant references.