How Web search works: delegated research and evidence

When a task needs public evidence, the coordinating agent delegates a bounded question to a dedicated Web research executor. An executor is a specialist agent that handles one delegated unit of work. It runs a short, source-led review in its own context: it searches, reads, compares, and reports before the coordinating agent continues the main task.

How a delegated Web research run reaches a task resultThe coordinating agent keeps the main task conversation and document history. It sends only a bounded research brief to a dedicated Web research executor. That executor creates focused queries, reviews candidate pages, reads selected sources, and returns a compact evidence package. The coordinating agent uses that package in the final task result.

Delegate one research question

The coordinating agent decides whether the task needs external evidence. For a broad or unclear subject, it first asks for a small evidence brief before it splits the work further. For a focused subject, it can delegate the research directly.

The router starts a new executor for that brief. The new context contains the brief, not the parent conversation, document history, or tool results. The executor has only Web search, page reading, and completion tools. It cannot call your registered tools or delegate the work again.

This separation keeps the research question apart from the wider task strategy. It also keeps full page content inside the research context: the coordinating agent receives the condensed evidence package, not every raw search result and page.

Run a bounded source review

The Web research executor turns the brief into an evidence checklist and several focused queries. It batches independent queries instead of sending one result back to the coordinating agent at a time. Search returns candidate pages ordered for relevance. The executor then opens the most useful pages and works from their extracted content and final URLs, not from search extracts alone.

Within the same delegated run, the executor can refine its queries, compare publishers, reject weak or duplicate sources, and record evidence gaps. The loop is bounded: it stops when the evidence is sufficient, when new searches add no material information, or when the research limit is reached.

Return evidence, then finish the task

The executor returns one structured package to the coordinating agent. It contains a concise answer, individual claims, sources that it reviewed, claim-linked citations with excerpts, and unresolved gaps. The evidence contract requires the executor to mark each cited source as read; it cannot use a search extract alone as a citation.

The coordinating agent combines this package with the rest of the task. Your terminal SDK result still follows the response model that your application requested. The SDK does not add the internal research package as a separate field, so request source fields in the final model when your application must retain them.

This process gives the coordinating agent several reviewed sources and explicit gaps. It does not make every source independent or every claim true. Request Web research shows how to retain source fields. Sources and factuality explains how to review them.

Understand the data boundary

Only the delegated brief enters the Web research context. Parts of that brief can become public search queries. Include only the detail needed to find the evidence, and do not include secrets.

Data protection covers query and page-request paths. Page access covers supported page behavior. Troubleshoot Web search covers missing or unreadable sources.