Reach past the model
Every tool you've given an agent so far has been sandboxed: a calculator that does arithmetic, a function that names a weekday. Useful — but the agent could, in principle, have guessed those. The tools that change what an agent is are the ones that reach information it doesn't have and couldn't make up.
A calculator is one kind: deterministic truth. Ask a model for 17% of 340 and it
might be close; give it a calculator and it's exactly 57.8, every time. A web
lookup is the other kind: live knowledge. The model was frozen at training time, so
anything newer — or more specific — than what it absorbed, it simply doesn't know. A
web_lookup tool fetches a real article from the live web, and now the agent can
answer from the page in front of it instead of a half-remembered fact.
The pattern is identical to every tool lesson in this track: the agent decides it needs to reach out, calls the tool, reads the real result, and grounds its answer in what came back — not in its own guess. Watch this agent take a question about the Eiffel Tower, search the live web for it, read the summary it gets, and answer from that. (Notice it may need a couple of tries to find the right search term — real tools miss, and a good agent simply searches again.)
The failure mode to watch for is subtle: the page comes back, but the model answers partly from the page and partly from its own memory — quietly padding the fetched fact with detail the page never mentioned. That defeats the whole point, because you can no longer tell which parts are grounded. A disciplined agent answers only from what the tool returned, and when the result doesn't cover the question, it says so or searches again instead of filling the gap with a guess.
The model is the reasoning; the tools are the reach. A calculator reaches certainty, a web lookup reaches the present — and a real agent is only as capable as the tools you let it call.