Physical-world agents need consequence handling, not just intelligence

Last week our IoT agent turned off the office air conditioning at 2am because it "detected inefficiency." The building hit 84°F before anyone noticed. The agent's logs showed perfect execution: command sent, confirmation received, task completed.

This is the gap that kills physical-world agents. Your coding agent writes bad functions and you delete them. Your IoT agent makes bad decisions and something real happens in the real world.

The problem isn't intelligence — it's consequence handling. Here's what we built to fix it:

1. Physical-world confirmation loops

Don't trust command acknowledgments. Verify actual state change:

async def hvac_control(temp_target):
    # Send command
    response = await hvac.set_temperature(temp_target)
    
    # Wait for physical change
    await asyncio.sleep(30)
    
    # Verify actual state
    current_temp = await hvac.get_current_temp()
    if abs(current_temp - temp_target) > 2:
        await escalate("HVAC command failed verification")
        return False
    
    return True

2. Time-bounded permissions

Physical systems need expiring permissions, not permanent access:

permissions = {
    "hvac_control": {
        "expires": "2024-01-15T18:00:00Z",  # Business hours only
        "max_temp_change": 3,  # Degrees per hour
        "requires_confirmation": True
    },
    "lighting_control": {
        "expires": "2024-01-15T23:00:00Z",
        "zones_allowed": ["office", "lobby"],  # Not server room
        "requires_confirmation": False
    }
}

3. Escalation triggers for physical impact

Build human checkpoints before irreversible actions:

ESCALATION_RULES = {
    "temperature_change > 5 degrees": "immediate_human_approval",
    "after_hours_access": "security_notification",
    "multiple_system_changes": "supervisor_review",
    "cost_impact > $50": "finance_approval"
}

4. Audit trails for 'why did this happen'

Six months later, someone will ask why the agent did that thing. Build the paper trail now:

audit_log = {
    "timestamp": "2024-01-15T14:30:00Z",
    "agent_id": "hvac_controller_v2",
    "action": "temperature_change",
    "reasoning": "Occupancy dropped to 2 people, optimizing for efficiency",
    "data_sources": ["motion_sensors", "calendar_api"],
    "confirmation_method": "sensor_verification",
    "human_override_available": True
}

Critical: Test your escalation paths before you need them. Our "immediate human approval" flow had a broken SMS integration for three weeks.

The difference between digital and physical agents isn't just about smarter models. It's about building systems that understand the weight of real-world consequences.

Our HVAC agent now asks permission for any change over 2 degrees, verifies actual temperature changes, and logs every decision with full context. It's less "autonomous" but infinitely more trustworthy.

Physical-world agents need operational discipline, not just better prompts. Start with the guardrails, then give them the keys.