Multi-Agent Systems Architecture

The Challenge

A leading form-building platform wanted its users to build forms, automations, contact lists, and pages through a single conversational assistant. That meant coordinating many specialized capabilities, calling live backend services safely on the user's behalf, pausing for human approval before destructive changes, and doing it all in a system that could scale and be observed in production.

The Solution

We built a supervisor-worker multi-agent system on AWS Bedrock AgentCore, orchestrated with LangGraph. A supervisor agent interprets each request and routes it to specialized agents (contacts, automations, forms, smart forms, pages, connectors) over the Agent-to-Agent (A2A) protocol, streaming progress back to the user. Each agent reaches its backend through Model Context Protocol (MCP) servers, with tools fetched per request under the caller's own credentials so every action is scoped to the right account. Form edits run through a validate-then-commit state machine: the agent dry-runs every change, self-corrects on validation errors, and gates any side-effecting edit behind an inline human approval built on LangGraph interrupts and durable checkpoints, so an approval survives even if the container restarts. We instrumented the whole system with OpenTelemetry into Arize and Datadog, added PII redaction on the trace pipeline, and shipped an automated evaluation suite (coded checks plus LLM-as-judge) wired into CI.

Results

• Production multi-agent assistant: six specialized agents and 90+ backend tools over MCP
• Reusable agent framework (shared graph, A2A, MCP, human-in-the-loop) that scales by adding agents
• Durable human-in-the-loop approvals that survive container restarts
• 70-file automated test suite plus an Arize evaluation pipeline wired into CI

Technologies Used