Skip to content

Agentic Workflows Phase 1 — LangGraph + MCP Foundation (consolidates Sprint 15 P1)

Agentic Workflows Phase 1 — Core LangGraph + MCP Foundation

Context

This is the core differentiator for GENIE.AI as the GovStack AI Building Block. Phase 1 establishes the foundation: LangGraph workflow engine, MCP proxy infrastructure, and initial workflow implementations.

This issue consolidates and replaces the Sprint 15 Phase 1 child issues (#269-#275) with a refined implementation plan.

Architecture (from Confluence Agentic Spec)

User Request → Backend BFF → LangGraph Orchestrator
                                    ├─ Tool: RAG Query (existing)
                                    ├─ Tool: MCP Server → GovStack BB API
                                    ├─ Tool: Translation
                                    └─ Tool: Guardrail (optional)
                                  
MCP Servers (standalone processes):
  ├─ Auth MCP (Security BB)
  ├─ Payments MCP (Payments BB)
  ├─ Scheduler MCP (Scheduler BB)
  └─ Registration MCP (Registration BB)

Scope

Phase 1A: LangGraph Core

  • Implement genie-ai-overlay/workflows/ package with LangGraph
  • Define WorkflowState Pydantic model (typed state passed between nodes)
  • Implement core node types: LLMNode, ToolNode, ConditionalRouter, HumanInputNode
  • Create workflow definition schema (JSON/YAML → LangGraph graph)
  • Integrate with existing ChatQnA as the entry point
  • Support both streaming (SSE) and non-streaming execution

Phase 1B: MCP Infrastructure

  • Deploy mcpo (MCP proxy orchestrator) as a Docker service
  • Implement MCP client in Python that calls MCP servers via stdio/SSE
  • Register MCP servers in the backend Node.js API
  • Implement WebSocket endpoint for real-time workflow status (#274)
  • Add Node.js API routes for MCP proxying (#273)

Phase 1C: Initial MCP Servers

  • Auth MCP Server: Authenticate user, verify permissions, get user profile (mock for now)
  • Payments MCP Server: Check payment status, initiate payment, get receipt (mock for now)
  • Scheduler MCP Server: Check availability, book appointment, cancel appointment (mock for now)

Phase 1D: Frontend Workflow Visualization

  • Basic workflow status component (shows current step, pending steps, completed steps)
  • WebSocket integration for real-time status updates
  • Error/retry UI for failed workflow steps

Phase 1E: Pause/Resume

  • Implement checkpointing in LangGraph (#275)
  • Persist workflow state to ArangoDB
  • Resume interrupted workflows from last checkpoint

Acceptance Criteria

  • LangGraph orchestrator handles multi-step workflows with typed state
  • 3 MCP servers deployed and callable from LangGraph nodes
  • Workflow status visible in frontend via WebSocket
  • Pause/resume works for interrupted workflows
  • Existing RAG chat flow still works (backward compatible)
  • Integration tests cover a sample 3-step workflow

Related Issues

  • Consolidates Sprint 15 Phase 1: #269, #270, #271, #272, #273, #274, #275
  • Depends on: Sprint 22 (streaming, auth), Sprint 23 (observability)
  • Phase 2 (Sprint 25): BPMN visualization, RAG integration, hybrid workflows, security/audit

Labels

1.Parent_Issue, ENHANCEMENT