Portal Agent Architecture — Decision Flow

System Graph

flowchart TD
    subgraph INPUT["📥 Case Intake"]
        A[Case arrives in portal] --> B{Source?}
        B -->|Portal scrape| C["🤖 Read Agent<br/><em>pulls case details,<br/>member info, clinical docs</em>"]
        B -->|Manual entry| D[User enters case ID<br/>in dashboard]
        B -->|Batch queue| E["Job queue feeds<br/>cases sequentially"]
    end

    subgraph AI["🧠 AI Review"]
        C --> F
        D --> F
        E --> F
        F["AI Clinical Review<br/><em>prior-auth-review skill</em>"] --> G{Determination}
        G -->|Approve| H["✅ Approve<br/>+ rationale + criteria cited"]
        G -->|Deny| I["❌ Deny<br/>+ rationale + criteria cited"]
        G -->|Need more info| J["⚠️ Request Info<br/>+ what's missing"]
        G -->|Uncertain| K["🔍 Human Review<br/>AI confidence too low"]
    end

    subgraph VERIFY1["👁️ Screen 1: Review Determination"]
        H --> L
        I --> L
        J --> L
        L["Dashboard shows:<br/>• AI decision + confidence<br/>• Clinical criteria matched<br/>• Supporting evidence<br/>• Rationale text"]
        L --> M{User approves?}
        M -->|Yes| N[Proceed to submission]
        M -->|Edit| O[User modifies decision/rationale]
        M -->|Reject| P[Back to manual review]
        O --> N
        K --> P
    end

    subgraph RPA["🖥️ Browser Agent"]
        N --> Q["Launch headless Chrome"]
        Q --> R["Step 1: Navigate to portal login"]
        R --> S["Step 2: Enter credentials"]
        S --> T{Login success?}
        T -->|No| U["❌ Auth failed<br/>alert user"]
        T -->|Yes| V["Step 3: Find case in portal"]
        V --> W{Case found?}
        W -->|No| X["Fallback: direct URL nav"]
        X --> W2{Found now?}
        W2 -->|No| Y["❌ Case not found<br/>alert user"]
        W2 -->|Yes| Z
        W -->|Yes| Z["Step 4: Verify case detail"]
        Z --> AA["Step 5: Upload attachments<br/><em>(if any)</em>"]
        AA --> BB["Step 6: Select decision + submit"]
        BB --> CC{Submission success?}
        CC -->|No| DD["Retry up to 2x"]
        DD -->|Still fails| EE["❌ Submission failed<br/>screenshots captured"]
        CC -->|Yes| FF["Step 7: Extract confirmation #"]
    end

    subgraph VERIFY2["👁️ Screen 2: Live Agent View"]
        direction LR
        R -.->|screenshot| LV["Browser viewport<br/><em>live screenshots<br/>per step</em>"]
        S -.->|screenshot| LV
        V -.->|screenshot| LV
        Z -.->|screenshot| LV
        BB -.->|screenshot| LV
        FF -.->|screenshot| LV
    end

    subgraph VERIFY3["👁️ Screen 3: Job Result"]
        FF --> GG["Dashboard shows:<br/>• ✅ Confirmation number<br/>• Step-by-step log<br/>• Duration per step<br/>• Click any step → screenshot<br/>• Full Playwright trace (debug)"]
        EE --> HH["Dashboard shows:<br/>• ❌ Error + which step failed<br/>• Screenshot at failure point<br/>• Playwright trace for debugging"]
        U --> HH
        Y --> HH
    end

    subgraph EVIDENCE["📁 Evidence Package"]
        GG --> II["Saved to disk:<br/>• result.json<br/>• screenshot per step (.png)<br/>• trace.zip (Playwright)<br/>• Kept for 7 days / 20 jobs"]
    end

    style INPUT fill:#1a1a2e,stroke:#4a4a6a,color:#e0e0e0
    style AI fill:#1a2e1a,stroke:#4a6a4a,color:#e0e0e0
    style VERIFY1 fill:#2e2e1a,stroke:#6a6a4a,color:#e0e0e0
    style RPA fill:#1a1a2e,stroke:#4a4a6a,color:#e0e0e0
    style VERIFY2 fill:#2e2e1a,stroke:#6a6a4a,color:#e0e0e0
    style VERIFY3 fill:#2e2e1a,stroke:#6a6a4a,color:#e0e0e0
    style EVIDENCE fill:#1a2e2e,stroke:#4a6a6a,color:#e0e0e0

What Exists vs. What's Needed

User Decision Points

The Pipeline (What "Useful" Looks Like)

Portal → Read Agent → AI Review → User Approval → Submit Agent → Confirmation
  ↑                                                                    ↓
  └──────────────── Evidence package saved for audit ──────────────────┘

The missing piece that makes everything click: Screen 1 (Review Determination). This is where the nurse sees the AI's work, agrees or edits, and triggers submission. Without it, the AI and the agent are two separate tools.