Multi-Agent LangGraph + MCP: Supervisor Pattern, Checkpoints & Human Approval

A single agent with every tool is a bottleneck. Research, billing lookups, code changes, and ops tickets want different prompts, models, and permissions. The LangGraph supervisor pattern puts a routing agent in front of specialist workers; MCP gives those workers a standard tool bus; LangChain still binds the LLM. Checkpoints and human-in-the-loop nodes are what make the graph safe to resume after a crash or an approval wait.

This is the production follow-up to a one-loop agent: how to split work, persist state, gate risky tools, and connect specialists to NestJS SaaS APIs without sharing a god-mode key.

When one agent is no longer enough

• The system prompt tries to be researcher, coder, and SRE at once — quality drops
• Write tools (refunds, deploys, emails) sit next to read-only search in the same allowlist
• Long tasks die mid-run because nothing persisted graph state
• You cannot pause for a human without losing the conversation

Supervisor + specialists + MCP

1. Supervisor — reads the user goal, picks a worker (or finishes). No direct write tools if you can avoid it
2. Specialists — e.g. Research (RAG/search MCP), Ops (logs/tickets MCP), Billing (read-only finance MCP)
3. MCP servers — one server per domain so tools, auth, and rate limits stay isolated
4. LangChain — tool schemas, structured output, and model adapters inside each worker node
5. LangGraph — edges for handoff, retry, interrupt, and END

Handoff is a graph edge, not a chat mention

Do not hope the supervisor "remembers" to call Alice. Model handoff as explicit LangGraph edges with a typed state object: goal, tenant id, last worker, artifacts, and pending approval. Workers return structured results; the supervisor only decides the next node.

Checkpoints: resume after crash or approval

• Persist LangGraph state (thread id + checkpoint) in Postgres or Redis — memory dies with the pod
• Human-in-the-loop is an interrupt node: graph waits, user/admin approves, execution continues from checkpoint
• Never re-run a write tool on resume without idempotency keys
• Store tenant and actor on the state so a resumed job cannot hop tenants

If your agent cannot pause for a human and resume without replaying a refund, you do not have a workflow — you have a lucky script.

MCP server design for specialists

• Split servers by blast radius: read-search vs write-billing vs deploy
• Pass the user/service identity into MCP; tools enforce RBAC the same way the NestJS API does
• Keep tool results small and structured — dump entire tables and you will overflow context
• Log every MCP invocation with correlation id from the graph thread

Production checklist

• Budget per worker and per whole graph (steps, tokens, wall clock)
• Eval suites per specialist — a billing worker regression should not hide behind a research win
• Dead-letter threads that exceed retries; operators inspect, do not auto-loop forever
• Feature-flag new workers; supervisor routing table is product config
• Trace supervisor decisions so support can explain why Ops was chosen over Research

What "done" looks like

A supervisor routes a ticket to Ops, Ops calls MCP, a refund waits on human approval, the graph resumes from a checkpoint, and audit shows who approved. That is multi-agent AI enterprises can run. Explore Cyber Infoware products or contact us for agent architecture, MCP design, or a SaaS backend specialists can trust.