A specific capability handed to an agent — web search, "read this file," "query the CRM," "send this email" — and one use of it. Between tool calls, the model reads the result and decides what to do next.
Why it matters: an agent's tool list is its job description and its risk profile. Reviewing what an agent actually did means reading its tool calls — the actions — not just its confident summary of itself.
More in “Giving an agent hands”