Agentic AI: From LangGraph Prototype to Enterprise Deployment

The gap between a working LangGraph prototype and a production multi-agent system is wider than most teams expect. We've crossed that gap a number of times now, and the patterns are consistent enough to be worth documenting.

What Breaks at Scale

Prototypes hide latency. A single-agent loop that completes in 2 seconds locally might fan out to 12 parallel tool calls in production, each with its own retry logic and timeout behavior. Your orchestration layer needs to treat latency as a first-class concern, not an afterthought.

State management is the other common failure point. LangGraph's checkpointing is excellent, but you need to design your state schema for failure recovery from day one. Ask: if this graph resumes from checkpoint mid-run, what invariants must hold?

Human-in-the-Loop is Non-Negotiable in Finance

For any agent that can initiate transactions, place orders, or modify records, human approval gates are required — not just good practice. We implement interrupt nodes at defined decision boundaries, with clear UI surfaces for human reviewers. The agent proposes; the human approves.

Deployment Architecture

We run agent workloads on Cloud Run with min-instances=1 for latency-sensitive flows and min-instances=0 for batch. Separate services for orchestrator and worker agents. Redis for shared state between parallel branches. Structured logging with trace IDs that span the entire graph execution.

The prototype proved the concept. The production system proved the discipline.