@pmcollab/coworkstream-engine-langgraph

Coding agents: read packages/workstream-engine/AGENTS.md for the engine integration recipe before wiring this adapter. Migrating an existing setup? See packages/workstream-engine/MIGRATION.md.

LangGraph (LangChain.js) AgentRunner adapter for @pmcollab/coworkstream-engine. Wraps a compiled graph as a workstream agent.

Install

npm install @pmcollab/coworkstream-engine-langgraph @langchain/langgraph

Use

import { StateGraph, MessagesAnnotation } from '@langchain/langgraph'
import { ChatAnthropic } from '@langchain/anthropic'
import { createLangGraphAgent } from '@pmcollab/coworkstream-engine-langgraph'

const llm = new ChatAnthropic({ model: 'claude-sonnet-4-6' })

const graph = new StateGraph(MessagesAnnotation)
  .addNode('reason', async (state) => ({ messages: [await llm.invoke(state.messages)] }))
  .addEdge('__start__', 'reason')
  .addEdge('reason', '__end__')
  .compile()

const agent = createLangGraphAgent({
  id: 'reasoner',
  graph,
  toInput: (item) => ({
    messages: [{ role: 'user', content: `Decide on: ${item.title}` }],
  }),
  extractPosition: (state) => {
    const last = state.messages[state.messages.length - 1]?.content ?? ''
    return {
      position: /approve/i.test(last) ? 'approve' : 'reject',
      reasoning: last,
      confidence: 0.7,
    }
  },
})

Why

LangGraph is the durable, stateful agent runtime in the LangChain ecosystem. Use this adapter when:

• You're already using LangGraph for multi-step agent workflows
• You want checkpointing/persistence of agent state
• Your agents call multiple tools / sub-graphs before reaching a decision

Options

createLangGraphAgent({
  id, graph,                    // required
  name?, avatar?, capabilities?,
  toInput?: (item, ctx) => unknown,
  extractPosition?: (state, item, ctx) => { position, reasoning, confidence },
  config?: unknown,             // passed to graph.invoke as 2nd arg
  defaultConfidence?: number,
})

License

Commercial. See LICENSE in the repository root.