Session Description

Traditional RAG assumes that retrieving a handful of relevant chunks provides sufficient context for an LLM to complete a task. Enterprise AI breaks this assumption. Organizations often have billions of files spread across complex folder hierarchies, multiple document versions, and long-form content. The challenge is no longer fitting more tokens into a model’s context window—it is progressively reducing an enormous search space into the small, high-quality evidence set needed for each reasoning step. Simply retrieving more chunks or leveraging larger context windows often introduces conflicting or irrelevant information, degrading reasoning rather than improving it.
In this session, we present Box’s Recursive Global Agent, a retrieval-first architecture that treats search as an iterative reasoning problem instead of a one-time preprocessing step. Rather than trying to fetch all relevant documents upfront, the agent recursively decomposes complex tasks into smaller subtasks while simultaneously partitioning the search space. Independent sub-agents perform targeted hybrid retrieval, extract only the evidence required for their specific task, and recursively merge compressed results into progressively richer context for downstream reasoning. This enables the system to analyze corpora far larger than any model context window while maintaining high retrieval precision and efficient parallel execution.
Using real enterprise workflows—including large-scale contract analysis, compliance reviews, and multi-document investigations—we’ll demonstrate how recursive agentic retrieval significantly improves context quality over traditional RAG approaches. We’ll also discuss trajectory-aware evaluation techniques that measure retrieval quality and reasoning across the entire execution path, enabling continuous optimization of agent behavior. Attendees will leave with practical design patterns for recursive query decomposition, hybrid retrieval, evidence compression, and evaluation-driven development that enable production AI agents to solve complex enterprise search problems beyond the limits of conventional RAG.