When people say "AI agent", the first question shouldn't be "what does it do?" but "what is it forbidden from doing?". In Nour, each agent has a narrow, written role — and its limits are guaranteed by how it's built, not just by instructions. Here's a tour of all three.
Its role: the first eye on every inbound message. It classifies each one as news (a new report or event), a reply (follow-up inside an existing conversation), or chatter (greeting, prayer, general question, engagement).
How it saves money: before asking the AI at all, a fast filter catches the obvious cases: lone emoji, short greetings, and any story reply or reaction — instantly marked "not important" at zero cost. Since ~93% of inbound traffic is exactly that, this filter alone eliminates most of the running cost.
Its role: scoring reports (only). It produces three things: a confidence score from 0–100, a gap list (source? place? time? visual evidence? eyewitness?), and a corroboration count — how many other reports in the last 24 hours describe the same event.
The most important thing about it is what it lacks: this agent never says "confirmed" or "unconfirmed". There is no green button that closes the case. The platform's entire vocabulary is "attribution & confidence", not "verification" — because judging a story's truth is a human editorial responsibility, and AI's job is to prepare, not to rule.
Its role: filling the gaps the attribution agent identified, by talking to the source directly. Its questions are simple and direct, always leading with the journalistically vital one: "if you have a photo or video, please send it" — then time, place, and whether they witnessed it themselves.
Why it's the most constrained agent: because it's the only one speaking to real people on the organisation's behalf. That's why its messages are fixed templates, not free-form LLM text — making it structurally impossible for it to:
One agent that "does everything" is an agent whose limits you can't guarantee. Three agents, each with a narrow role and limits built into the code — that's a system you can audit and trust.
And that's what makes Nour fit for a newsroom: not just its intelligence — its discipline.