All case studies
Case study 01AI SystemsFull-StackAutomation

Multi-Agent Client Support Platform

LangGraph agents that route, delegate, and act on business systems safely through MCP.

Role
Architect and lead engineer
Core stack
LangGraphLangChainFastMCPPostgreSQLLangSmith
−28%
Inference cost from context budgets and model routing
p95
Latency and tool failures monitored in production
5
Engineers mentored through architecture and code reviews

Overview

A production agent platform for client support workflows: a supervisor agent routes requests to specialist agents, which act on business APIs through MCP, with human approval for sensitive actions.

Support workflows span several business systems and include actions that must never run twice or without the right permissions. The agents had to be reliable across retries and long-running conversations, pause for human approval, and stay affordable at scale.

Product walkthrough

The flow, screen by screen

Agent console
01 / 03

Agent console. The supervisor routes a billing request to the specialist agent. The run trace shows each MCP tool call, its latency, and the checkpoint where the run paused for approval.

ArchitectureCorePlatformAsync / control
tool callsstatejobsloggedLangSmithevals · tracingAudit trailtool calls · approvalsClient channelssupport UI · APIFastAPI gatewayauth · streamingSupervisor agentLangGraph routingSpecialist agentsdelegated tasksHuman approvalresumableRedis workersasync runsPostgreSQLtyped state · checkpointsMCP serverFastMCP · scoped authBusiness APIsidempotent actions
Request flow: the gateway streams to the supervisor, which delegates to specialists; tools run through the MCP server with approvals, checkpoints, and audit logging.

Approach

  1. 01Architected LangGraph agents with supervisor routing and specialist delegation, using typed state so every step has a well-defined contract.
  2. 02Persisted state in PostgreSQL checkpoints with bounded retries, so runs resume cleanly after failures or human approvals.
  3. 03Introduced MCP integrations with FastMCP for business APIs, enforcing validated tool schemas, scoped authorization, audit trails, and idempotency keys.
  4. 04Delivered asynchronous agent execution and streaming React/FastAPI interfaces on Redis, Docker, AWS, and Terraform.
  5. 05Established LangSmith regression suites for task completion, tool selection, and grounding, alongside failure-injection tests.

Key decisions

Supervisor + specialists over one large agent

Smaller, task-scoped agents are easier to evaluate, prompt, and permission than a single agent with every tool.

Idempotency keys on every side-effecting tool

Retries are inevitable in distributed systems; idempotency keys make them safe instead of producing duplicate actions.

Resumable human approvals

Checkpointed state lets a run pause for a person and continue exactly where it stopped, without holding resources open.

Outcomes

  • Validated schemas and scoped authorization keep agents within their permissions.
  • Audit trails record tool calls and approvals for review.
  • LangGraph
  • LangChain
  • FastMCP
  • PostgreSQL
  • LangSmith
  • FastAPI
  • React
  • Redis
  • Docker
  • AWS
  • Terraform

Working on something similar?

I work with clients worldwide, across any time zone. Tell me what you're building.

Send a message Email meMore case studies