How does Box leverage AI for its customers?

Box uses AI to extract insights from vast amounts of unstructured content stored on its platform, enabling customers, from large banks to small firms, to derive value from their data.

What is MCP and how does it help scale AI?

MCP is a framework that allows for configuring connectors to integrate Box's AI capabilities into an agent, enabling users to leverage Box's data without writing custom integration code.

What are the benefits of using A2A for scaling AI?

A2A allows for the creation of multiple, purpose-built agents orchestrated by a central component. This modular approach enables agents to perform single tasks efficiently and allows for easy addition of new capabilities.

How does the live demo showcase A2A's capabilities?

The demo shows three agents (orchestrator, files, and extraction) working together to extract key data from invoices stored in Box, demonstrating how A2A can automate repetitive tasks in an enterprise setting.

What are the key architectural patterns of A2A?

A2A offers pluggable tools and models, a single integration point to simplify complex API interactions, and the ability to separate concerns by assigning specific tasks to individual agents.

What are the security and governance implications of enterprise AI like A2A?

Security is paramount. Robust access control ensures users only see data they're permitted to, and Box performs security checks before AI processing to prevent unauthorized access to sensitive information.

How do developers choose between different AI agent SDKs like A2A, OpenAI, and Grock?

The choice often comes down to personal preference and the existing technology stack. Developers should consider what works best within their environment rather than switching frameworks solely based on a presentation.

AI Dev 25 x NYC | Scott Hurrey: Scaling Enterprise AI with MCP and A2A

DeepLearning.AI

Education5 min read28 min video

Dec 5, 2025|458 views|5

Save to Pod

Key Moments

TL;DR

Scaling enterprise AI: Box uses MCP and A2A for modular, secure AI integrations.

Key Insights

Box handles 90% of enterprise data as unstructured content, making AI crucial for insights.

MCP simplifies AI-to-tool connections by standardizing integrations, reducing custom code.

A2A enables modular, multi-agent systems where each agent has a specific purpose.

The demo showcased A2A's ability to process invoices by orchestrating file finding, data extraction, and reporting agents.

Key architectural patterns for A2A include pluggable tools, single integration points (MCP), separation of concerns, and robust security.

Security and governance are paramount, with Box enforcing access control before AI processes any data.

THE CHALLENGE OF UNSTRUCTURED ENTERPRISE DATA

Scott Hurrey from Box highlights the significant challenge enterprises face with unstructured data, which constitutes 90% of their content. Unlike structured data easily queried in databases, unstructured content like documents and presentations historically required manual review to extract insights. Box, as a cloud content storage engine, stores over an exabyte of data for more than 120,000 customers. The advent of generative AI presents a powerful opportunity to unlock value from this vast, underutilized data, moving beyond simple chatbots to sophisticated AI agents capable of automating complex tasks.

BOX'S APPROACH TO AI INTEGRATION: MCP AND EXTERNAL SYSTEMS

Box offers its own AI engine and tools like Box Hubs for managing vector stores and AI agents for composing text and answering questions. However, they recognize that enterprises use numerous systems beyond Box. To address this, Box focuses on building partnerships and enabling seamless data integration. They provide resources like a cookbook for Grok integration and have worked with Pinecone to aggregate data from multiple sources. Langchain libraries for Python and JavaScript are also available, allowing developers to access Box content within their Langchain applications.

SIMPLIFYING AI-TO-TOOL CONNECTIONS WITH MCP

The Model Context Protocol (MCP) is presented as a solution to simplify AI-to-tool connections, eliminating the need for custom integration code. Instead of writing complex scripts to navigate APIs, extract data, and handle variations in document formats, developers can configure an MCP connector. This allows AI agents to access Box functionalities directly. Box offers both an open-source MCP server, actively used for demos and feature development, and a remote, enterprise-grade MCP product that adheres to strict engineering best practices and provides essential security and compliance features. The MCP server handles permission management, ensuring that agents only access data permitted by the user's or service account's tokens.

ENABLING MODULARITY WITH AGENT-TO-AGENT (A2A) PROTOCOL

The Agent-to-Agent (A2A) protocol facilitates the creation of modular, multi-agent systems designed for scalability. In this architecture, multiple purpose-built agents, each responsible for a single, well-defined task, work together. An orchestrator agent manages these specialized agents, determining which agent performs which task and in what sequence. This modular approach allows for complex problems to be solved efficiently. New agents can be added to the system without impacting existing ones, and the orchestrator can dynamically discover and utilize these agents through the A2A protocol, minimizing the need for glue code.

LIVE DEMONSTRATION: INVOICE PROCESSING WITH A2A

A live demo illustrated the practical application of A2A for invoice processing. Three agents were configured: an orchestrator, a files agent to locate invoices in Box, and an extraction agent to pull key data like client name, invoice amount, and product name using Box's MCP server tools. The demo successfully extracted information from multiple invoices, showcasing the agents' ability to work in sequence. The output was a JSON object containing the extracted data, demonstrating A2A's potential for automating repetitive enterprise tasks and integrating with reporting or document generation systems.

ARCHITECTURAL PATTERNS AND SCALABILITY THROUGH A2A

A2A offers several key architectural benefits for scaling AI. It allows for pluggable tools and models, enabling easy integration of new AI capabilities or MCP server updates. MCP provides a single, simplified integration point, abstracting complex API calls into single tools for agents. This approach supports the separation of concerns, where agents are specialized and maintainable by dedicated teams, reducing interdependencies. Furthermore, A2A, combined with Box's built-in security and governance, ensures that agents only access authorized data through token-based access control, crucial for highly regulated industries. The concept of scaling with tools like Autonomy, where agents can be spun up in parallel for individual files, was also highlighted as a method for significantly speeding up processing times.

SECURITY, GOVERNANCE, AND BEST PRACTICES

Security and governance are central to enterprise AI, especially within Box's ecosystem. Box enforces strict access controls, verifying user or service account permissions before any data is processed by AI. This ensures that sensitive information is never exposed inappropriately. When building MCP servers, it's recommended to design tools based on what agents need to accomplish rather than directly mapping API specifications. Limiting an agent to only the MCP tools required for its specific task is also crucial to prevent confusion and ensure efficient operation. Regular evaluation of AI models against specific content sets helps in selecting the right model for the right use case and building more reliable agents.

FUTURE OF ENTERPRISE AI AND DEVELOPER RESOURCES

The discussion points towards autonomous AI agents as the future, forming a workforce capable of making decisions and handling complex operations. Box is committed to empowering developers to build these advanced systems. They offer resources such as courses on deep learning platforms, which delve into these concepts with hands-on exercises, and plan to release follow-up content for advanced topics. The session emphasized that while various SDKs for agent development exist, the choice often comes down to personal preference and the existing technology stack within an organization, with A2A and MCP offering a robust framework for enterprise-grade solutions.

Mentioned in This Episode

●Products

●Software & Apps

●Companies

●Organizations

●Concepts

Data Structure of Enterprise Information

Data extracted from this episode

Data Type	Percentage
Structured Data	10%
Unstructured Data	90%

Common Questions

The primary challenge is that 90% of enterprise data is unstructured (like Word documents and PowerPoints), making it difficult to extract insights compared to the 10% structured data.

Topics

A2A Data Extraction Unstructured Data AI Scalability LLM Integration

Mentioned in this video

conceptEnterprise-grade security and compliance

A core value proposition of Box, maintained by the MCP server, ensuring that users only access files they are permitted to see through token and permission management.

softwareOrchestrator agent

The agent that sits on top of other purpose-built agents in an A2A architecture, deciding which agent performs which task and in what order.

conceptOCR

Optical Character Recognition, a process needed for scanned PDFs to convert them into a format usable by LLMs.

softwareStrand's SDK

An SDK for building AI agents, mentioned in a question about comparing different agent frameworks.

softwareQuestion and Answer Agent

An agent provided by Box that allows users to ask questions and receive answers.

softwareOpenAI agents SDK

An SDK from OpenAI for building AI agents, discussed as an alternative to A2A.

productBox Hubs

A feature within Box that allows users to organize up to 20,000 files in a hub and automatically maintains the vector store.