How it works

When Web Search is enabled on a prompt version, the agent’s model can decide to call the tool with a query string. The tool runs the search against a DuckDuckGo-backed index and returns a clean, structured list of results that the model can summarise, cite, or follow up on by fetching the full page with the companion Web Fetch tool.

Each result is a small, transparent object with three fields:

• title — the page title as indexed
• href — the destination URL (citation-ready)
• body — a short snippet, typically one to three sentences

The configuration surface is deliberately tiny. There is a single optional field, max_results, which caps how many results come back; left unset, the tool returns the first batch. There are no API keys to manage and no per-project quotas to provision, because the DuckDuckGo backing means there is nothing to set up. Typical responses complete in one to three seconds.

One design choice worth calling out: returned snippets are treated as untrusted input by default. Downstream prompts are not meant to interpret them as instructions, which hardens agents against the classic “a poisoned search result hijacks the model” attack. That posture is built into the tool’s guidance rather than something you have to remember to add.