An agent is one published rule that runs every evening against every liquid stock on the exchange and reports what it found. Put as many on watch as you like; your desk is what they report back.
Two kinds, and the difference is on every card. A rule agent is deterministic Python — the rule is published, and you can reproduce its output from the same prices. A model agent is one where a language model genuinely runs, and both of them are fenced in: the filing reader discards any figure whose quote is not verbatim in the filing, and the note writer discards any sentence in which a new numeral appeared. Calling the other forty-seven “AI” would be marketing rather than description.
No agent tells you what to buy. Each one reports which stocks met a stated condition, and what that condition has been worth historically — including when the answer is nothing.