A tool call is a model asking an external system to run something and waiting for the answer before it continues. The model picks a tool by name, fills in arguments that match the tool's declared schema, and receives a structured result. Everything an agent does beyond generating text happens through one.
Why it matters
A tool call is the smallest thing you can meter, log, price, permission or debug. When an agent workflow goes wrong, the useful question is almost never "what did the model think" and almost always "which tools did it call, with what arguments, and what came back".
It is also where cost lives. A server that charges per result charges per call, so the shape of your tool schema is a pricing decision as much as an engineering one.
Example
You ask an agent for UK fintechs that raised a Series A recently. It makes one tool call:
search_signals({
industry: "fintech",
country: "GB",
fundingStage: "series_a",
fundedWithinDays: 60,
limit: 20
})The model wrote none of that JSON by hand. It read the tool's schema, mapped your sentence onto the parameters that exist, and called it. Then it read the result and wrote you a sentence back.
Common mistake
Assuming a failed tool call fails loudly. Many do not. A tool that returns zero results because a filter value was rejected looks identical to a genuinely empty market, and the model will report the empty market with total confidence.
Good tool design closes that gap by returning a distinct error rather than an empty list. Signl's search_signals returns an unknown_industry error listing the values it does accept, precisely so that a zero-result answer means something.
How this relates to Signl
Signl exposes five tools: search_signals, get_weekly_drop, get_credit_balance, enrich_lead and refresh_person_signals. The first three are free. Credits are charged on enrich_lead only, and only when it returns a verified contact, so a call that finds nothing costs you nothing.