Rather than a frontier model issuing its own search tool calls and reading raw results inline, it hands a query to a specialized subagent that decomposes the query into subqueries, issues searches, inspects and reads sources, and returns a distilled evidence package. This frees the calling model's context window and inference budget for reasoning rather than search mechanics. Mixedbread's Toast 1, launched in August 2026, is an example — it works with any existing search backend via an OpenAI-compatible API and is one of several specialized entrants in this category, alongside models like SID-1 and Chroma's Context-1.