Dubai mobility, answered with its sources.
Masar plans how to answer each question, gathers evidence across documents, the analytical database, geography and deterministic fare arithmetic — then grades that evidence and re-plans if it falls short. Every claim carries a citation you can open.
A real recorded runMULTI_HOP
A1A2A3understand the question
A4planned 4 sub-tasksA8
A12 · insufficient — 2 named gapsback to the Planner
A4re-planned · 5 sub-tasksA6A8
A12 · insufficient — 1 named gapback to the Planner
A4re-planned · 5 sub-tasksA8A6
A12 · cycle cap reachedA13 answers · low confidence · 2 citations
The Grader judged the evidence insufficient twice and sent it back to re-plan. At the cycle cap it answered with low confidence rather than inventing facts — that restraint is the point.
Local fallback model, ~152s. Cloud providers answer in seconds; the path is identical.
Ask the last one to see how it handles a question the data genuinely cannot answer.
Agent trace
Ask something and the fourteen agents will light up here as they run.