Box goes AI native by building its agent on an open framework

Box rebuilt Box AI, the agent it ships to more than 100,000 enterprise customers, on LangGraph Deep Agents, and the engineering write up is unusually candid about the architecture. The earlier design used hardcoded specialist sub agents, a dedicated search agent, a QA agent and a compose agent, which added latency on simple requests. The new design uses a single parent Global Agent that classifies intent at runtime and dynamically spawns child agents only when a task actually needs them. A representative workflow: pull every contract from the past ten years above a value threshold and score each against a risk rubric. The parent produces a plan and fans out, one child searching documents, another retrieving the rubric in parallel, and a third synthesizing once the first two finish, each running in an isolated context window. Two hard requirements drove the framework choice: complete model agnosticism, since Box lets customers pick among OpenAI, Anthropic, Google and others, and speed of iteration. Box says the open harness delivered about 3x faster iteration by letting its engineers focus on enterprise specific problems instead of rebuilding core agent plumbing (company claim). The lesson for a smaller team is that the build path no longer means writing orchestration from scratch. You adopt a framework for the agent loop and spend your engineering budget on your own domain.