Box 在 BoxWorks 发布 Box Architect、Context Retrieval、Box…

In just the past few years, AI models have gotten unbelievably good at reasoning, logic, multimodal capabilities, tool use, computer use, and deeper domain expertise across nearly every high-value area of knowledge work. And inside the AI labs, there is now the view that we are reaching a form of recursive self-improvement, where AI itself is helping discover and drive research breakthroughs in model training, which means that you get faster and faster progress with every generation.

Combined, this means that we can finally deploy AI at some of the hardest problems in enterprises. We are no longer in a period where living at the frontier is just having a conversation with an AI assistant and getting a quick answer back, but instead being able to deploy AI agents across highly complex domains in an enterprise and letting the agents go off and work as long as they need to complete the task. As a result, AI agents will increasingly be deployed behind the scenes in nearly every workflow in an enterprise, including areas like automating client onboarding, accelerating drug discovery pipelines, reducing risk in an enterprise by reading every critical contract, analyzing research data to accelerate product development, and much more.

Now, to do this, AI agents need to know everything about the business they're being used in. They are inherently trained on either the public internet or privately acquired data sources. What they don't know about are the unique aspects of your organization. A company's unique product roadmap, financial decisions that were made in a planning process, contractual terms that they signed with clients, research breakthroughs that are happening somewhere in the organization, and all of the other elements that are living inside of our enterprise data. Critically, that information lives inside of our unstructured data, which represents 90% of the corporate information that companies create, store, and work with.

For companies to stay competitive in the 21st century, they will need to figure out how they can uniquely apply this context, combined with their people and processes, with these new superintelligent AI systems. We are truly at the dawn of a new era of how industries will operate, and the way we manage and work with our information sits at the very center of this.

Today, at BoxWorks, we're making 4 major announcements around the Box platform to change how enterprises work with their unstructured data and AI:

1) Box Architect, a new way to autonomously build out workflows and management content

2) Box Context Retrieval, a new platform capability designed in partnership with NVIDIA to give agents exactly the context they need to do important work

3) Box Mount, a new way to give agents a file system to securely work with, in addition to all new MCP capabilities to enable any AI agent to work with enterprise content

4) All new threat detection and security capabilities to help protect the flow of data in the enterprise with agents

Starting with Box Architect. AI is not just about helping us be more productive as individuals, but it can truly change how we run and operate the ongoing workflows in our enterprises. Given the rate of improvements we are seeing with frontier AI, we are now confident that AI can help improve an organization's processes and workflows far faster than was ever possible before. What if we could take the same frontier intelligence that AI labs are using to develop their own models, drive math and bio breakthroughs, or code the most advanced software in the world, and use this frontier AI to help improve our customers' workflows in Box?

Most businesses are held back by their access to the talent and resources to help automate and streamline their workflows. We have far more processes inside of companies than the people and time to help automate that work. And because AI is trained on the broadest sets of knowledge and domain expertise, what if you could actually use AI to not just automate your existing processes, but fundamentally redesign them and reinvent them in a way far better than how they are run today? And taking this even further, what if AI could actually help tune and optimize those workflows and processes over time, just as an industry or technical expert might within your organization?

This is Box Architect. With Box Architect, enterprises get the full power of frontier AI to autonomously build, test, optimize, and run their content-centric processes. Box Architect effectively acts as an agentic operator of Box's core content and workflow capabilities -- like Box Apps, Automate, Forms, Doc Gen, Agents, and more -- to help customers design any content agentic workflow in Box. Along with major updates to our Box Agent to handle much longer-running and more complex work, new Box Automate features for advanced workflows, new Extract optimization capabilities, and Box Apps components to power more comprehensive no-code apps in Box, Architect will help accelerate our customers' transformation of their content workflows with AI.

Next up, we're announcing a new partnership with NVIDIA to help power advanced Context Retrieval in Box. Box's new Context Retrieval platform capability delivers hybrid search functionality which combines keyword and semantic search to deliver better results for agents, and respects Box user permissions so agents only retrieve what the user is allowed to access. Internal testing shows our new Context Retrieval functionality delivers a 20% improvement in accuracy, a 90% increase in token efficiency, and 4X faster answers, giving customers a secure and governed way to connect AI (both Box Agents, as well as external agents via MCP, CLI, or our direct APIs) to enterprise content that is more accurate, faster, and more cost-effective.

Coming soon, Box Mount will let AI agents work directly with Box folders in secure computing environments in their sandbox, without the need to manually download, move, or upload files. Agents will be able to read, edit, and create files while keeping Box permissions, security controls, and file syncing in place. Box Mount will also integrate with NVIDIA OpenShell for secure agent operations, DigitalOcean for managed agent sessions, as well as other leading sandbox providers like Daytona, E2B, AWS, Docker, and more. We're also expanding our MCP capabilities and support within ChatGPT, Copilot, Claude, Grok, Slack, Atlassian, Gemini, Legora, Cursor, Harvey, and any other platform with Box MCP support.

Finally, major hacks and rogue agents show the reason why enterprises must better protect their crown jewels. We're doubling down on Box Shield to secure content against emerging cyber threats, AI-accelerated attacks, and agent sprawl. New Box Threat Detection and Box Agent Security enhancements will detect anomalous downloads, sessions, and agent activity; enable adaptive responses like download restrictions, step-up verification, and session termination; and introduce agent security controls including classification-based access, MCP guardrails, threshold-based alerts, and audit trails of agent activity.

We're just at the beginning of a new era of how we work with our information, and at Box, we are building the platform to help companies take full advantage of this moment.