Summary
Scott Hurrey, Director of Developer Relations at Box, demonstrated at AI Dev 25 NYC how enterprises can combine MCP (Model Context Protocol) and Google’s A2A (Agent-to-Agent) protocol to build scalable, multi-agent document processing pipelines. With Box managing over an exabyte of content across 120,000 customers—drawn from IDC data showing 90% of enterprise data is unstructured—the motivation is concrete: AI agents need structured, governed access to Word documents, PowerPoints, and contracts that traditional query systems cannot reach.
Hurrey’s live demo featured a three-agent system: an orchestrator agent that dynamically discovers and delegates to specialized sub-agents via A2A’s agent card protocol; a files agent that queries Box for relevant invoices from a specified folder; and an extraction agent that uses the Box MCP server to pull key-value pairs—client name, invoice amount, product name—from each document. Dynamic agent discovery means new agents can join the network without rewriting orchestration glue code, a significant advantage for scaling enterprise deployments.
The talk also covers same-day partnership announcements, including a new cookbook integrating Box’s MCP server with Groq’s Responses API, plus existing integrations with Pinecone, LangChain (Python and newly released JavaScript), and Workday. Hurrey closes with a practical warning: limit each agent to only the MCP tools it actually needs, as tool overload is a documented cause of agent confusion in systems like Claude Code and Cursor.
📺 Source: DeepLearningAI · Published December 05, 2025
🏷️ Format: Hands On Build
