WRU

WRU ("where are you") splits family errands between AI agents without anyone sharing coordinates. To answer "who's picking up the kid?", many families turn on always-on location sharing, even though all they need is who can make it in time. In WRU each family member has a Muse agent that sees only that person's coarse status (home / near home / on the way) and a rough ETA, read from their own self-hosted location server. The agents negotiate in rounds with offer / counter / commit / refuse messages, a parent approves the result once, and each phone receives only its own task. Privacy is enforced, not promised. Each agent holds a token that can read only its owner (anyone else's status returns 403). Agents' own words are never published: reasons become fixed templates and ETAs appear only as ranges. We measured leakage with a real Muse Spark run: 50 families across Seoul's 25 districts, with Muse playing both the agents and the attacker. With the gate on, the 50 families produce just 9 distinct logs; the best-case attacker finds the home district 28% of the time and the Muse attacker 4% (chance). Without the gate, every log is unique and the home district is found 100% of the time. Built during the hackathon: the negotiation engine, Muse clients, privacy gate, leak test, demo UI and token boundary. WhereCtx, the location server WRU reads from, is my separate pre-existing project and not part of this entry.