Hunt 3-bed houses under $3.5m in Bondi, Bellevue Hill, Randwick, Coogee & Vaucluse — must walk to a station, decent school catchment, and not be overpriced vs comps. Buying to live in, ready in ~3 months.
The brief sits at the top as editable instructions — tap any underlined token to tune budget, suburbs or must-haves. Below it, a board of parallel tasks (Cursor/Codex "mission control"): each run shows live status, an honest tool trace, and a Steer / Run now / Stop control. You delegate work, then supervise N streams.
Frames the agent as staff you direct, not a chatbot you query. The funnel (214 → 26 → 8) makes its judgement legible at a glance, and a finished run hands back a reviewable artifact (top find + verdict) rather than prose.
Every computed dollar figure is tagged blue and traces to its comps ("↳ 6 sold comps · SEIFA decile 9"). Verdict chips (under/over) are computed; the run's narration is the agent's read. Tool names (computeCMA()) keep the work honest.
The middle take. More ambitious than the inbox (A) — parallel runs, editable brief, intervene mid-run — but still a calm board, not the full immersive theatre of Variant C.
Real: suburbs, the funnel logic, the data sources, the steerable-brief concept. Faked: addresses, dollars, timers and progress are illustrative; images are picsum placeholders.