%20-%20AI%20Multi-Agent%20Planning%20System%20.jpg&w=3840&q=75)
Arc - AI Multi-Agent Planning System (Product Management AI)
A LangGraph multi-agent planning system that turns a goal into research, analysis, critique, and an executable plan. WebSocket streaming, checkpoints, and persisted state make long-running sessions visible and resumable.
A LangGraph multi-agent planning system that turns a goal into research, analysis, critique, and an executable plan. WebSocket streaming, checkpoints, and persisted state make long-running sessions visible and resumable.
What made this system difficult?
Needed realtime visibility into multi-step agent runs, plus reliability features like checkpoint/resume and consistent persistence across reconnects.
How did the design respond?
Implemented StateGraph workflows and a FastAPI WebSocket endpoint that validates project ownership and streams workflow events/state sync. Added token/cost tracking backed by a pricing table and a checkpoint-driven resume flow for long-running executions.
How responsibility moves through the system.
Coordinator Agent
- — Workflow orchestration
- — Task routing & coordination
Research Agents
- — Web research & synthesis
- — Structured evidence gathering
Planner & Critic
- — Structured plan generation
- — Review/critique loop
Overview
- Solo-built LangGraph-based multi-agent planning system (Arc/PMAI).
Purpose & User Flow
Purpose
- Automate planning: take a goal and produce structured deliverables via agent division of labor + critique loops.
User Flow
- Input goal/context → run workflow (agent routing) → stream step-level progress and artifacts (WebSocket) → resume from checkpoints → export deliverables.
Key Components
- StateGraph workflows with conditional routing between agent roles (
backend/agents/graph.py) - FastAPI WebSocket streaming endpoint with state sync/resume (
backend/api/websocket.py) - Token/cost tracking with pricing table + provider balances (
backend/agents/token_tracker.py) - Checkpoint-based resume flow (
resume_workflow)