Agentic Workflow
The agentic workflow converts an incoming request into a structured, reviewable, and executable operational plan.
Trigger Points
POST /requests: primary trigger for request-driven planning.POST /resources/suggest/{disaster_id}: resource-focused suggestion trigger.
Workflow Pipeline
flowchart TD
T1[Trigger received] --> T2[Context assembly]
T2 --> T3[Request + disaster + resource retrieval]
T3 --> T4[LLM planning and reasoning]
T4 --> T5[Task decomposition]
T5 --> T6[Resource and manpower suggestion]
T6 --> T7[Confidence and priority scoring]
T7 --> T8[Persist WorkflowOutput]
T8 --> T9[Human review/moderation]
T9 -->|approved| T10[Create tasks and assignments]
T9 -->|rejected/edited| T11[Revise plan]
T11 --> T8
Data Artifacts Produced
WorkflowOutput(per request)WorkflowTask[](ordered response steps with priority and approval status)WorkflowResource[](resource recommendations with quantities and approval status)
Moderation And Human-in-the-Loop
- AI output is not treated as final dispatch by default.
- Admin/dispatcher can approve, reject, or edit generated tasks/resources.
- Approved artifacts are translated into first-class operational entities (
Task, resource status updates).
Execution State Model
stateDiagram-v2
[*] --> Requested
Requested --> Planned: workflow generated
Planned --> AwaitingApproval: persisted for moderation
AwaitingApproval --> Approved: admin approves
AwaitingApproval --> Rejected: admin rejects
Rejected --> Planned: regenerated/edited
Approved --> Assigned: tasks created and assigned
Assigned --> InProgress: responder on_route
InProgress --> Completed: task success
InProgress --> Failed: operational failure
Observability and Reliability
- Each workflow run should be traceable (Langfuse integration).
- Failures in worker execution should not block API request acknowledgment.
- Retries, idempotent writes, and explicit status transitions reduce duplicate dispatch risk.