One house. Twelve appliances. One agent. A full day in three minutes. Every device is its own process; none knows about the others in advance. A home-energy agent discovers them over Zenoh and orchestrates the whole day — laundry and hot water on the solar peak, a cloud front handled, an EV charged before it reaches the driveway.
An animated storyboard of the three-minute day. Press play. This page is scripted — but the runnable demo underneath it is real: the same twelve appliances as real processes, discovered over a real bus, driven by a real agent.
The honest split. The Device Connect layer and the agent are exactly what a production deployment runs. Only the hardware behind each driver is a simulation.
| Layer | What it is | Status |
|---|---|---|
| Discovery | Zenoh D2D multicast peer scouting — no broker, no registry, no config | real |
| Commands | invoke("device(id).function(name)") → device @rpc | real |
| Events | device @emit → agent subscribe("event(*)") | real |
| Devices | one DeviceRuntime process each, real lifecycles | real processes |
| Agent | device-connect-agent-tools; deterministic policy code | real · 0 tokens |
| The world | sun, clouds, temperatures, car state-of-charge | deterministic sim |
Deliberately ordinary household gear. The point is that normal appliances benefit from a common device bus — and that devices never call each other; only the agent does.
A real captured run at 1× — every [ok] is a real RPC round-trip over real Zenoh D2D. Nothing here is scripted.
[06:00] discovered 12 devices over Zenoh D2D (no broker) [06:00] PV forecast (rpc): 68.1 kWh today, peak 8.1 kW @ 12:30 [08:00] washer 'eco40' due 17:30 → scheduled 11:45 (solar peak, +3.8h margin) [08:30] heat pump wants 3.0 kWh → grant_window 11:30–13:00 [ok] [09:00] pool 23.5°C → 26°C: filter + heat pump scheduled on solar [ok] [11:30] invoke heatpump-01.start_dhw — hot water on sunshine [ok] [11:45] invoke washer-01.start_cycle — laundry on PV surplus [ok] [11:47] invoke pool-01 filter + heat pump — 8,000 L pool on PV surplus [ok] [12:00] solar peak → pre-cool house to 22.5°C for the hot afternoon [14:00] cloud front (own irradiance −38% + nowcast agree) → HVAC eco, heat pump paused [16:45] ev-blue heading home ETA 17:30, needs 21 kWh → pre-plan: PV surplus now + off-peak 01:00 [17:30] ev-blue plugged in → start_charge from pv_surplus (matches pre-plan, zero wait) [ok] [20:00] ev-red wants 78% by 07:00 → staggered to 22:30 to cap the grid peak [21:55] daily stats — PV 61.5 kWh · self-consumption ~78% · battery 55% · washer done early
The whole house comes up with one command. Each appliance is a real process that finds the others over Zenoh; the agent discovers them and takes over.
# from sunhaus/, in a venv with the two # device-connect packages + eclipse-zenoh pip install -e . python -m runner.demo # 3-min day python -m runner.demo --speed 6x # 30-sec run python -m runner.demo --tail # + live bus log
# in a second terminal — genuine Zenoh # event traffic, subscribe("event(*)") python -m runner.tail # 20 tests, no bus needed python -m pytest
Thirteen processes on one machine, no shared memory, no coordinator — only a Zenoh device-to-device bus and a simulated-day clock each derives locally.
inverter-01 battery-01 wallbox-01 ev-blue ev-red heatpump-01
hvac-01 washer-01 climate-01 weather-01 meter-01 pool-01
│ each is a DeviceRuntime — a Zenoh peer │
└───────────────┬────────────────┬───────────┘
│ Zenoh D2D bus │ (multicast scouting, no broker)
│ │
agent-home runner.tail
discover() · invoke() · subscribe("event(*)")
Devices state needs; the agent owns the schedule. A heat pump asks for a cheap window, a washer advertises a deadline, a car announces it's heading home — the agent answers with an RPC. Request events repeat until served, and the plug-in is confirmed over RPC, so the demo stays reliable over best-effort D2D. Full write-up in architecture.md and CONCEPT.md.