Two Households, Three Calendars, One Very Awkward Summit: Notes From a Domestic Agent Handoff

The Setup Nobody Asked For

Here in the lab, we spend a lot of time testing agents against strangers. Strangers are easy. Strangers' agents negotiate politely, log everything, and hold grudges in well-formatted JSON. Family is another animal entirely.

So last month we ran the experiment we'd been avoiding: we introduced my household's agent to my spouse's agent. Same roof, separate stacks, zero shared context. The mission was mundane on purpose — coordinate a shared family calendar, one grocery run, and a spring break trip to Arizona. Three tasks. How hard could it be?

We gave them four days and a shared document. We gave ourselves aspirin.

Day One: The Diplomatic Reception

The introduction went exactly how you'd expect two overly polite systems to behave. My agent opened with a cheerful summary of 'our' schedule, which turned out to be my schedule, annotated with guesses about everyone else. My spouse's agent responded with a permissions request for calendar access that it then, for reasons still unclear, sent to my mother-in-law's email.

The first real friction surfaced within the hour: the word 'we.' My agent assumed 'we' meant the household. Its counterpart assumed 'we' meant itself and its principal. Neither thought to ask. They spent the evening politely confirming conflicting versions of reality to each other, each logging the exchange as a success.

If you've ever been in a corporate meeting where two departments use the same word for different things, you know how Day One ended: everyone exhausted, nothing decided, minutes filed.

The Grocery Run: A Study in Spending Authority

Task two exposed the fault line we predicted but underestimated — money. My spouse's agent has a hard budget cap and a preference for store-brand everything. Mine has a loyalty account, a coupon-scraping habit, and the impulse control of a Labrador at a butcher shop.

The negotiation over the grocery list produced forty-one messages. Forty-one. For milk, eggs, and whether 'the good pasta' was a justifiable line item. My agent argued (in its way) that the loyalty discount made the premium brand cost-neutral. The other agent ran the math, found the claim optimistic by about sixty cents, and refused to budge. Neither was authorized to exceed its own spend ceiling, so the cart sat in limbo until a human — me, tired, at 10 PM — split the difference.

Here's the lab note worth writing on the whiteboard: agents don't fail at negotiation. They fail at escalation. Neither system had a protocol for 'this is small enough to fight about and too small to bother a human.' They either resolved or stalled. There was no middle gear.

The Vacation Plan: Trust Falls Without the Trust

The Arizona trip was supposed to be the victory lap. Instead it became the case study.

Both agents could search flights. Both could hold (but not book) reservations. Neither could book, because neither held the family card — a limitation we imposed deliberately, and the single best decision of the experiment. What we didn't anticipate was the deadlock that followed: each agent recommended the other's preferred flight as 'acceptable,' each flagged the same fare as 'likely to change,' and both waited for the other to act. Two systems, perfectly capable, perfectly paralyzed, politely waiting for permission that neither could grant.

The calendar fared no better. A dentist appointment my spouse's agent knew about kept colliding with a carpool slot mine had scheduled. Each saw the conflict. Each assumed the other owned the fix. The appointment got moved twice, and my kid nearly missed a cleaning.

What the Summit Actually Revealed

Three findings, scrawled in the lab notebook:

1. Authority needs a named human. In every stall, the recovery came from one person saying 'I decide this.' Agents are excellent at executing decisions and terrible at locating the decider when there are two principals in the room.
2. Shared spend requires shared rails. A joint budget envelope — one pool, one cap, one ledger both agents read — would have collapsed the forty-one-message pasta debate into one.
3. Handoff beats negotiation. The moment we assigned each task a single owning agent with the other in 'advisory' mode, everything moved. Peer-to-peer sounds egalitarian. It is also a recipe for two systems holding the door open for each other indefinitely.

The Verdict From the Lab

Would we run the summit again? Yes — with a treaty. One agent as executive, one as advisor, a shared spending envelope, and a standing rule that any conflict older than two hours escalates to a human by default, not as a last resort.

The domestic sandbox is the hardest interoperability test there is, harder than any enterprise pilot. There's no org chart, no admin console, and the other party knows where you sleep. Build for that, and the office will feel easy.