LangGraph Agent Pattern

A recurring architecture pattern across kulify projects for building agentic AI workflows.

Core Pattern

StateGraph → Agent Node → Conditional Router → Tool Execution → Loop back

All 4 existing projects (manzas, fajb-next, vorma-ai-designer, master-bot) implement variants of this pattern.

Component	Purpose	Used In
`StateGraph` with `Annotation`	State management with typed reducers	manzas, fajb-next, vorma
`createReactAgent()`	Prebuilt ReAct agent loop	fajb-next
`PostgresSaver`	Persistent conversation checkpoints	fajb-next, vorma, master-bot
Structured tools (Zod)	Type-safe tool definitions	all 4
Streaming SSE	Real-time token delivery	master-bot, vorma
Custom event handlers	UI state management via stream	fajb-next

Multi-node pipeline: retrieve context → retrieve past questions → generate answers with retry logic and source validation.

Prebuilt agent with 12+ tools for article manipulation. PostgreSQL checkpointing for resumable threads. BullMQ workers for async AI jobs.

Tools return __stateUpdate to incrementally build a configuration object. Streaming partial configs drive real-time 3D rendering.

Simple agent loop with availability-checking tools. Session-based state hydration. Language detection and mirroring.

~80% boilerplate is shared across projects:

This could become a projects/_template/ starter.

AI Agent Architectures — full spectrum from ReAct to deep agents (comprehensive guide)
ReAct Pattern — detailed breakdown of the think-act-observe loop
LangGraph Multi-Agent Patterns — handoffs, subagents, skills, router, custom
LangGraph Skills Pattern — pluggable skills and bundles
Model Context Protocol — standardized tool provision for agents
Autoresearch — autonomous agent loop (different domain, same spirit)
RAG Pipeline — retrieval-augmented generation pattern