Box AI builds its enterprise content agent on LangChain's Deep Agents

Box, an enterprise content management platform with more than 100,000 customers worldwide, has moved its AI system to LangChain's Deep Agents architecture. The new Box Agent can search, analyze and synthesize information from corporate repositories on its own, while fully respecting the access rights already configured in the system.

From chatbot to agent

The difference between a regular AI chat and an agent is fundamental. A regular chat answers a question about a single document or a small excerpt. An agent plans a task, calls the right tools, refines its result iteratively and can work with thousands of files at once.

Box Agent is built on the Deep Agents framework from LangChain. This means the agent can:

• Break complex questions down into subtasks
• Call search and analysis tools in the right order
• Work with long contexts drawn from many documents
• Check and refine its conclusions iteratively
• Produce structured reports with links to sources

For an enterprise user, this means they can ask, "What is our travel expense policy, and how has it changed over the past two years?" and get a coherent answer assembled by the agent from dozens of documents, with no manual searching.

Security without compromises

The main problem of enterprise AI is not answer accuracy but respect for access rights. A large organization has hundreds of user groups with different levels of access to documents. If an AI agent ignores these restrictions, it becomes a serious security hole and a compliance violation.

Box solved this by integrating Box Agent directly into its existing permissions system. The agent sees only the files a given user can access at that moment. If a document is marked confidential or is available only to a specific department, the agent will not reveal its contents to anyone who is not entitled to see it.

This is critical for industries with strict security requirements: financial companies, law firms and healthcare organizations. These are the sectors where Box is most active, and where access-control requirements are the strictest.

Provider independence

Another key feature of Box Agent is complete independence from any specific AI provider. The Deep Agents architecture allows switching between different language models without reworking the agent's logic.

In practice, if one model is optimal today and a more powerful or cheaper one appears tomorrow, Box can update the backend without changing the user-facing product. With new AI models arriving every few weeks, this architectural flexibility becomes a significant competitive advantage.

LangChain acts here as an orchestration layer that separates the agent's business logic from specific LLM providers. Box can optimize model choice per task: one model for search, another for synthesis and generating the final report.

What this means

Box Agent is one of the first examples of a mature enterprise platform moving to a deep-agent architecture in production. This is neither a pilot nor a demo: Box serves some of the world's largest companies, and its bet on Deep Agents means agentic AI is ready for enterprise workloads with real security and scalability requirements.

For the rest of the market, it is a signal: the "agent plus enterprise access rights" pattern works and is becoming a working template for the industry.