How Box Built a Third-Generation AI Platform: Deep Agents with a Dynamic Parent/Child Model

Box, the enterprise content management platform with tens of thousands of organizational customers, has gone through three clearly distinct development phases in building its AI capabilities. The details of that evolution — and the architecture of the current system, called Deep Agents — were documented through a conversation with Box’s AI engineering team, published by the LangChain platform on June 12, 2026.

Three Generations of Box AI

Phase 1 — Single-document Q&A: The first implementation allowed users to ask questions within a single file. Useful for isolated analysis, but with no ability to synthesize information from different sources.

Phase 2 — Knowledge Hubs (RAG): The next step introduced a RAG-based layer that allowed queries against predefined document sets. This raised utility, but required sources to be defined in advance and could not dynamically understand complex inter-document relationships.

Phase 3 — Deep Agents: The current architecture builds on a full agentic paradigm. The system can analyze thousands of documents, understand complex contextual relationships, and produce synthesized conclusions that previous generations were unable to generate.

Why Is the Dynamic Parent/Child Structure Key?

The core of the Deep Agents architecture is a hierarchical model in which the Global Agent (parent) receives a request and classifies the user’s intent. Based on that classification, the parent dynamically creates child agents tailored to the specific task — at runtime, not in advance.

The key difference compared to alternative approaches: parent and child agents share the same tool registry without predefined specializations. If the request is simple, the parent can respond immediately without spawning a child agent. If the task is complex and multi-dimensional, the parent determines which children to create and which tools to assign them. Child agents work in isolated context windows, preventing cross-contamination of context and enabling cleaner parallel execution.

The tool registry encompasses BM25 keyword search, vector search, structured Q&A over tables, and file operations. Box’s team concluded that a single unified tool registry is more effective than dynamic sub-setting — agents always have full visibility into available capabilities.

Development Speed: from Three Months to Three Weeks

Box’s team reports that the parent/child architecture was delivered 4 times faster than the previous approach based on hard-coded specialized sub-agents. While the earlier design required three months of development, the new system was ready in about three weeks. The iteration cycle for new agents has been reduced to approximately two weeks, meaning Box can respond quickly to new business requirements without lengthy development cycles.

The team also notes 3× faster iteration thanks to the flexibility of the architecture itself — changes to delegation logic do not require rewriting hard-coded links between specialized agents.

Model-Agnostic Design and Context Management

Box built an abstraction layer completely insensitive to the choice of LLM provider. Clients can use OpenAI, Anthropic, and Google models without any changes to the business logic of the system.

Custom middleware addresses several key operational challenges:

• Parallel citation generation for streaming responses, without blocking the main flow
• Prompt caching for multi-turn conversations that reduces latency for repeated requests
• Automatic context summarization when conversation history exceeds 170,000 tokens — a threshold that ensures the model does not lose response coherence

Security and Permissions Without Compromise

One of the foundational requirements for an enterprise platform is respecting existing security models. The Deep Agents system does not bypass Box’s permission model — synthesis of information from thousands of documents occurs exclusively within the access rights of the user who submitted the request. Security integration was not bolted on afterward but was an architectural requirement from the beginning of development.

Box’s experience documents a trend increasingly visible across enterprise AI deployments: rigid, predefined agent networks deliver weaker results than dynamic hierarchies that make delegation decisions at runtime rather than at design time.