Glean Waldo: An agentic search model that pairs with LLMs to deliver frontier intelligence with ~50% lower latency and ~25% fewer tokens

Introduced Glean Waldo, a specialized agentic search model built on NVIDIA Nemotron 3 Nano, designed to optimize enterprise AI search latency and token usage. Waldo acts as an initial planning agent, executing tool plans (Glean Search, employee search, web search) before invoking a frontier model. This architecture reduces end-to-end latency by ~50% and token usage by ~25% compared to using a single frontier model for both planning and reasoning. Waldo is trained using Direct Preference Optimization (DPO) on production tool-use patterns and reinforcement learning with a task-specific reward signal based on retrieved document relevance. The post also details how Waldo's execution signals are used to enable adaptive reasoning, dynamically determining the required reasoning depth for a query, further contributing to efficiency.