Web search: delegated, source-led research for agent tasks
Web research in Duale AI delegates a bounded question to a dedicated executor that searches, reads, and synthesizes public pages.
Web research delegates a bounded question to a dedicated executor that searches public pages, reads selected content, and returns a source-linked synthesis to the coordinating agent.
- The coordinating agent delegates a bounded research question when a task needs public evidence.
- The executor starts with the delegated brief and cannot call registered tools or delegate work.
- Web search does not provide private-network access, non-HTTP protocols, or an interactive browser session.
- The brief can become a public search query, so include no secrets.
- Treat fetched text as external input that can affect later agent work.
Summaries were generated by AI. Generative AI is experimental.
Duale AI agents use a dedicated Web research executor to search and read public pages during a task. Use Web research when a task needs information that can change after model training, must compare public sources, or must find and read a public document.
Your application submits a normal task. The SDK has no separate Web search method. The coordinating agent decides when the task needs public evidence and delegates a bounded research question. The Web research executor turns that question into focused searches, reads selected pages, compares the material, and returns a compact source-linked synthesis to the coordinating agent.
What Web research includes
- Capability
- Delegate a public research topic
- What happens
- A bounded brief starts a separate Web research context
- Boundary
- The full task conversation and document history are not copied into it
- Capability
- Find public pages
- What happens
- The executor runs several focused queries and receives candidate pages
- Boundary
- A search result helps discovery; it is not source evidence
- Capability
- Read selected pages
- What happens
- The executor receives extracted content, a title, and the final redirected URL
- Boundary
- Page access is not an interactive browser or a supported sign-in flow
- Capability
- Compare source material
- What happens
- The executor returns claims, cited sources, excerpts, and unresolved gaps
- Boundary
- The final SDK result does not expose this research record as a separate field
Use precise terms
This documentation calls public-page discovery Web search. Web research is the larger workflow that delegates a question, runs several searches, reads selected pages, compares sources, and synthesizes the evidence. Sources and factuality defines grounding and fact-checking.
Web search is not general Internet access. It does not provide supported access to private-network destinations or non-HTTP protocols. It also does not expose clicks, forms, credential entry, screenshots, or an interactive browser session.
Keep research context narrow
The Web research executor starts with the delegated brief, not the coordinating agent’s conversation or document history. It cannot call your registered tools or delegate work to another agent. It can search, read public pages, and return its synthesis. The brief can become a public search query, so include only the detail the research needs and no secrets. Data protection defines the external paths and retention.
Treat fetched text as external input. The threat model explains how it can affect later agent work and where your integration must enforce approval.
Choose a guide
- How Web search works—follow a delegated source review from research brief to task result.
- Request Web research—ask for dated sources, gaps, and a typed result.
- Sources and factuality—decide what a source link proves and what you must verify.
- Page access—check which page behavior and session state the agent can use.
- Troubleshoot Web search—recover from missing, stale, blocked, or unreadable sources.
Scope and limits states permitted use. The Agent harness explains why the model, not your SDK code, uses these tools.