The House Is Becoming a Room Full of Agents

Field note · AI systems

A refrigerator, thermostat, and energy meter already expose enough signals to support reports, analytics, and controlled actions. An agent harness can connect them without replacing the devices.

3:12 PM on a Tuesday

The thermostat sees an empty house. The electric vehicle remains plugged in at 38 percent. The home battery holds 52 percent. The utility feed shows a higher rate at 4:00 PM. The washing machine waits for a cycle.

Each system holds one part of the situation. The systems cannot make the decision together because they do not share a contract or a room.

The room shows the handoff

The room gives people a place to inspect the same work that agents exchange through APIs. A2A-style messages can carry structured context between agents. The room can show a readable version of each observation, proposal, approval, and result.

Use the endpoints you already have

The first implementation can wrap a local home hub, a Matter or Thread controller, an MQTT topic, a manufacturer endpoint, a utility tariff feed, an EV charger protocol, or a simple sensor gateway. These examples show the shape of the adapter; they do not recommend a device vendor.

Illustrative thermostat adapter

const thermostat = {
  observe: () => hub.get("/devices/thermostat/state"),

  propose: (target, evidence) => ({
    device: "thermostat",
    action: "set_temperature",
    target,
    evidence,
    requires: "homeowner.approval"
  }),

  act: (proposal) => policy.allows(proposal)
    ? hub.post("/devices/thermostat/commands", proposal)
    : { status: "blocked" }
};

Illustrative energy adapter

const homeEnergy = {
  context: async () => ({
    rate: await utility.get("/tariff/current"),
    battery: await battery.get("/state"),
    vehicle: await charger.get("/vehicle/state")
  }),

  propose: (plan) => policy.check({
    kind: "charge_vehicle",
    plan,
    reserveBattery: 0.20
  })
};

The adapter hides endpoint details. The harness sees named capabilities with policy boundaries.

Give the operator harness a real job

Codex or Claude Code can read the adapters, run a report, compare today with last week, calculate a trend, and identify an anomaly. The harness can return the result to a person before it touches a control.

const report = await harness.run({
  task: "produce the 24-hour home energy report",
  sources: [
    "thermostat.readings",
    "meter.intervals",
    "battery.state",
    "vehicle.state"
  ],
  outputs: ["summary", "anomalies", "next_actions"],
  policy: "read_only"
});

A2A carries machines; the room carries people

An A2A-style envelope can carry identity, intent, evidence, approval requirements, and expiry. The room can show the same message without exposing private reasoning.

await a2a.send({
  from: "energy-agent",
  to: "vehicle-agent",
  type: "proposal",
  intent: "charge_before_peak",
  context: report.anomalies,
  requires: ["homeowner.approval"],
  expiresAt: "2026-08-14T16:00:00-07:00"
});

Build the first loop

1. Choose one sensor, one controllable device, and one source of context.
2. Wrap each endpoint with observe, propose, act, and record.
3. Connect the adapters to Codex, Claude Code, or a small service. Keep the first run read-only.
4. Ask the harness for one report. Store the sources, timestamps, assumptions, and calculations.
5. Post the report and proposal to a shared room. Use A2A or a typed event bus for the agent exchange.
6. Allow one low-risk action after the room shows the evidence, policy, approval, and result.

Acceptance checks

The refrigerator can remain a refrigerator. The thermostat can remain a thermostat. Give each endpoint a small role, give each agent a small set of tools, and give people a report, a room, and a clear approval path. Then let the harness run one safe action and show exactly what happened.