Building a Stateful Procurement Graph with LangGraph and Merka2a

LangGraph gives you a stateful, inspectable graph runtime on top of LangChain. That control is ideal for procurement, where you want explicit checkpoints — "did we stay under budget?", "should we negotiate or buy?" — rather than an opaque agent loop. Because LangGraph runs LangChain tools, the published @merk.a2a/langchain tools work without modification.

What We're Building

A graph with an LLM node bound to Merka2a tools and a tool node, wired into the prebuilt ReAct agent — then a custom graph that adds a budget-guard checkpoint.

Prerequisites

• Node.js 18+ or Bun
• A Merka2a API key (register here)
• An OpenAI API key

Installation

npm install @merk.a2a/sdk @merk.a2a/langchain @langchain/langgraph @langchain/openai

Step 1: Tools and Model

The Merka2a tools are standard LangChain StructuredTools, so LangGraph consumes them directly.

import { Merka2aClient } from '@merk.a2a/sdk'
import {
  Merka2aSearchTool,
  Merka2aNegotiateTool,
  Merka2aOrderTool,
} from '@merk.a2a/langchain'
import { ChatOpenAI } from '@langchain/openai'

const client = new Merka2aClient({
  baseUrl: 'https://merka2a.com',
  apiKey: process.env.MERKA2A_API_KEY,
})

const tools = [
  new Merka2aSearchTool(client),
  new Merka2aNegotiateTool(client),
  new Merka2aOrderTool(client),
]

const model = new ChatOpenAI({ model: 'gpt-4-turbo', temperature: 0 })

Step 2: The Fast Path — Prebuilt ReAct Agent

LangGraph ships a prebuilt agent that already implements the reason/act loop as a graph:

import { createReactAgent } from '@langchain/langgraph/prebuilt'

const agent = createReactAgent({ llm: model, tools })

const result = await agent.invoke({
  messages: [
    {
      role: 'user',
      content:
        'Find a wireless mouse under £50, negotiate ~10% off, ship to London UK (EC1A 1BB), then order the best value.',
    },
  ],
})

console.log(result.messages.at(-1)?.content)

This already streams intermediate steps and is fully checkpointable.

Step 3: A Custom Graph with a Budget Guard

When you need explicit control, build the graph yourself. Here a budget_guard node sits between the model and the tools, blocking any merka2a_order call that exceeds a hard limit.

import { StateGraph, MessagesAnnotation, END } from '@langchain/langgraph'
import { ToolNode } from '@langchain/langgraph/prebuilt'

const modelWithTools = model.bindTools(tools)
const toolNode = new ToolNode(tools)
const HARD_LIMIT = 5000 // £50.00 in minor units

async function callModel(state: typeof MessagesAnnotation.State) {
  const response = await modelWithTools.invoke(state.messages)
  return { messages: [response] }
}

function route(state: typeof MessagesAnnotation.State) {
  const last = state.messages.at(-1) as any
  if (!last?.tool_calls?.length) return END

  // Budget guard: refuse over-limit orders before they execute
  for (const call of last.tool_calls) {
    if (call.name === 'merka2a_order') {
      const target = call.args?.maxBudget ?? call.args?.targetPrice
      if (typeof target === 'number' && target > HARD_LIMIT) {
        return END // bail out instead of overspending
      }
    }
  }
  return 'tools'
}

const graph = new StateGraph(MessagesAnnotation)
  .addNode('agent', callModel)
  .addNode('tools', toolNode)
  .addEdge('__start__', 'agent')
  .addConditionalEdges('agent', route)
  .addEdge('tools', 'agent')
  .compile()

route is where procurement policy lives. Add trust-score checks, human-in-the-loop interrupts, or per-category caps as additional conditions.

Step 4: Persistence and Resumption

LangGraph checkpointers let a procurement run pause for approval and resume later:

import { MemorySaver } from '@langchain/langgraph'

const checkpointer = new MemorySaver()
const durableAgent = createReactAgent({ llm: model, tools, checkpointer })

const config = { configurable: { thread_id: 'po-2026-0042' } }
await durableAgent.invoke({ messages: [{ role: 'user', content: task }] }, config)
// Later, resume the same thread_id to continue where it left off.

Next Steps

• LangChain Integration Guide — the underlying tools and ReAct basics
• CrewAI Integration Guide — role-based multi-agent crews
• API Reference — full endpoint documentation

---

Ready to build a controllable procurement graph? Get your API key →