Review work

After an agent takes action on a Business report, Middesk stores a record of its findings and how each finding was applied or not applied to the report. This record is represented as an agent_application in the API.

Understanding the results

Each agent_application answers three questions:

• What did the agent find? resolutions contains one entry for each reviewed finding. original_artifact preserves what the agent proposed and any supporting sources, while agent_display_name identifies the agent that produced it.
• What was done with those findings? outcome records how the finding was handled. resolved_artifact, correction, and exclusion show any adjustments. actions lists the actions taken, and execution_result shows whether they were applied, partially applied, left the report unchanged, or failed.
• How did the Business report change? changes contains the effects recorded for completed actions, and report_transitions summarizes relevant before-and-after report values. At the application level, review_task_transitions records corresponding changes to the Business’s review tasks.

Checking screenshot capture status

When web evidence screenshot capture is enabled for your agents, the screenshot_captures object tells you whether captures have completed for an agent run.

Only one state indicates pending work, so check for that state rather than for a timestamp:

const captures = application.screenshot_captures;
const screenshotCapturePending =
  captures !== null && captures.settled_at === null;

screenshot_captures reflects the status for the result’s run_id. If another agent run applies findings, its new result may show pending captures even when earlier results are settled.

When a screenshot is available, its source includes a screenshot_id that you can use to retrieve the Document.

If your workflow requires available screenshots in the Business PDF, retrieve the latest PDF when no captures are pending.

Receiving updates through webhooks

Middesk sends a business.updated event after an agent finishes acting on its findings and its agent_application can be reviewed. If screenshots are still pending, Middesk sends another business.updated event after all capture operations finish.

Default webhook payloads don’t include agent_applications. To include these records and their screenshot status, add agent_applications and the nested screenshot_captures field to a custom event schema.

Use the webhook event schema API to configure these fields, then choose which webhook events to subscribe to and when they are sent.

Agent application fields

Select only the fields your integration uses. The following selection shows every available field:

{
  id
  agent_applications {
    object
    id
    run_id
    thread_id
    created_at
    finalized_at
    applied_by {
      id
      name
      email
    }
    screenshot_captures {
      settled_at
    }
    review_task_transitions {
      key
      before
      after
    }
    resolutions {
      object
      id
      artifact_id
      agent
      agent_display_name
      original_artifact
      resolved_artifact
      report_record {
        object
        id
      }
      artifact_field
      outcome
      decision_mode
      proposed_value
      resolved_value
      correction {
        id
        original_value
        resolved_value
        reason
        created_at
      }
      exclusion {
        id
        created_at
      }
      selected_sources
      execution_result
      actions {
        id
        type
        status
        completed_at
      }
      changes {
        kind
        report_record {
          object
          id
          namespace
        }
        operation
        before
        after
      }
      report_transitions {
        key
        label
        kind
        before
        after
        before_intent
        after_intent
      }
    }
  }
}

Each result contains these top-level fields:

Automating how findings are applied

Each item in resolutions includes decision_mode, which shows whether its outcome was chosen manually or automatically. Learn about configuring your interrupt policy to control which findings apply automatically and which pause for human review.