LangGraph and MCP: enterprise agent deployment overview

Harish Gaggar describes an enterprise AI agent framework at Intuit Credit Karma, covering the infrastructure around a LangGraph workflow. An internal portal creates a repository with memory, tracking, evaluation, PII filtering and context optimization components. It also provisions supporting resources such as workload identity and secret management. Local development leads into container builds, two PR approvals, security scanning and Kubernetes deployment.

A centralized MCP gateway authenticates requests and routes them to reviewed tools on custom MCP servers. Gaggar recommends building internal servers for air-gapped environments. The workflow classifies operations, gathers evidence and checks policies; sensitive writes require human approval, with decisions recorded in audit logs.

The talk separates gateway monitoring from agent tracking. OpenTelemetry collects traces and metrics, while agent records capture model use, token costs and tool execution. Threads, runs and checkpoints preserve workflow state, with SQLite used in the local demonstration.

The PII layer aims to intercept secrets before they reach models or stored records. Its effectiveness remains a control to evaluate. A context optimizer prunes, reranks, deduplicates and compresses inputs. Gaggar reports a reduction from 24,000 to 2,400 tokens for one query, but the demonstration does not establish equivalent answer quality or measured business ROI.