An agent that searches the web

A model knows the world as of its training. An exchange rate, a timetable, a competitor’s price, yesterday’s news: it knows none of those and, worse, will cheerfully make them up. Here it has a search tool and an instruction to show where the answer came from.

What is inside

1. Customer question — the entry.
2. Agent with search decides for itself whether to search. A simple question it answers directly; a question about today sends it to the web, possibly more than once.
3. Web search is wired with link_extension, which makes it a tool in the model’s hands rather than a step in the chain. It returns page addresses and excerpts, not just a summary.
4. Answer with links — the exit.

The links appear because the agent’s system prompt asks for them. Remove that line and you still get an answer, with no way to check it.

What to set up for yourself

• How many results (max_results on the search) and how many searches per turn (max_searches on the agent). Each search is billed separately from the model call.
• Domain limits. include_domains and exclude_domains narrow the search to sources you trust. For a bot that talks about law or prices, that is the difference between an answer and a rumour.
• The system prompt. Keep the requirement to cite; add what to do when a search returns nothing.

What to connect

Nothing. Search runs on the platform and is paid for in quarks from the balance; no separate key is needed. Do remember a searching turn costs more than a plain one: you pay for the search and for the tokens the excerpts take up on their way into the model.