Cambridge Research: Agentic AI for Advanced HVAC Building Control

The Building Energy Problem

Commercial buildings account for approximately 40 percent of global energy consumption, and HVAC systems represent the single largest component of that footprint, typically consuming 50 to 60 percent of a building's total energy. Despite decades of building management system (BMS) development, most commercial HVAC systems still operate on static schedules and simplistic rule-based control logic. Thermostats follow fixed setpoints. Ventilation runs at constant rates during occupied hours regardless of actual occupancy. Chillers and boilers cycle based on outdoor air temperature thresholds established during commissioning and rarely updated.

The result is massive waste. The International Energy Agency estimates that intelligent control systems could reduce HVAC energy consumption by 20 to 40 percent in existing commercial buildings without any hardware modifications. The challenge has been developing control systems sophisticated enough to balance energy efficiency, occupant comfort, equipment longevity, and cost optimization simultaneously in real time.

Researchers at Cambridge University's Department of Engineering have published a framework that addresses this challenge using an agentic AI approach they call Office-in-the-Loop.

The Office-in-the-Loop Framework

The Cambridge research introduces a multi-agent system where specialized AI agents collaborate to manage different aspects of building climate control. Unlike centralized optimization approaches that treat the building as a single control problem, Office-in-the-Loop decomposes the challenge into distinct agent roles, each with its own perception, reasoning, and action capabilities.

flowchart LR
    CALLER(["Homeowner"])
    subgraph TEL["Telephony"]
        SIP["Twilio SIP and PSTN"]
    end
    subgraph BRAIN["Field Service AI Agent"]
        STT["Streaming STT<br/>Deepgram or Whisper"]
        NLU{"Intent and<br/>Entity Extraction"}
        TOOLS["Tool Calls"]
        TTS["Streaming TTS<br/>ElevenLabs or Rime"]
    end
    subgraph DATA["Live Data Plane"]
        CRM[("CRM and Notes")]
        CAL[("Calendar and<br/>Schedule")]
        KB[("Knowledge Base<br/>and Policies")]
    end
    subgraph OUT["Outcomes"]
        O1(["Service appointment booked"])
        O2(["Quote sent via SMS"])
        O3(["Tech dispatched today"])
    end
    CALLER --> SIP --> STT --> NLU
    NLU -->|Lookup| TOOLS
    TOOLS <--> CRM
    TOOLS <--> CAL
    TOOLS <--> KB
    NLU --> TTS --> SIP --> CALLER
    NLU -->|Resolved| O1
    NLU -->|Schedule| O2
    NLU -->|Escalate| O3
    style CALLER fill:#f1f5f9,stroke:#64748b,color:#0f172a
    style NLU fill:#4f46e5,stroke:#4338ca,color:#fff
    style O1 fill:#059669,stroke:#047857,color:#fff
    style O2 fill:#0ea5e9,stroke:#0369a1,color:#fff
    style O3 fill:#f59e0b,stroke:#d97706,color:#1f2937

Agent Architecture

The framework deploys four primary agent types:

• Zone Comfort Agent: Monitors temperature, humidity, CO2 levels, and occupant feedback in each building zone. This agent maintains a dynamic comfort model for each zone that adapts based on actual occupant preferences rather than static ASHRAE standards
• Energy Optimization Agent: Tracks real-time electricity pricing, solar generation output, battery storage levels, and grid demand signals. It continuously calculates the most cost-effective way to deliver the thermal comfort that zone agents request
• Equipment Health Agent: Monitors compressor cycles, fan motor current draw, filter pressure differentials, and refrigerant levels. This agent adjusts operating parameters to extend equipment life and flags maintenance needs before failures occur
• Coordination Agent: Arbitrates between the competing objectives of the other three agents. When the comfort agent requests maximum cooling but the energy agent identifies a peak pricing period, the coordination agent negotiates a compromise that keeps comfort within acceptable bounds while minimizing cost

How Agents Perceive and Act

Each agent maintains its own sensor data streams and builds internal models of its domain:

• Occupancy sensing via a combination of CO2 concentration analysis, WiFi device counts, and calendar integration provides real-time room-by-room occupancy without invasive surveillance
• Weather forecast integration from multiple meteorological APIs enables predictive pre-conditioning, cooling a building mass during off-peak hours ahead of a forecasted heat wave
• Energy price feeds from real-time wholesale markets and demand response program APIs inform cost-optimal scheduling
• Occupant feedback loops through simple mobile app interfaces allow building users to report comfort levels, training the zone agents on actual preferences

The agents communicate through a shared message bus, exchanging structured observations and negotiating actions through a defined protocol. This architecture prevents any single agent from making decisions that undermine another agent's objectives.

Research Results

The Cambridge team tested Office-in-the-Loop in a 12,000 square meter office complex over a six-month period spanning summer and winter conditions. The results demonstrated significant improvements across all measured dimensions.

Energy Savings

The agent-based system achieved a 28 percent reduction in total HVAC energy consumption compared to the building's previous BMS configuration. The savings came from three primary mechanisms:

• Occupancy-responsive ventilation reduced unnecessary air changes by 35 percent during partially occupied periods
• Predictive pre-conditioning shifted 40 percent of cooling load to off-peak electricity pricing windows
• Equipment optimization reduced compressor cycling by 22 percent through smoother load management

Thermal Comfort

Despite the significant energy reduction, occupant comfort actually improved. The system achieved thermal comfort satisfaction scores above 90 percent in post-occupancy surveys, up from 74 percent under the previous static control system. The improvement came primarily from eliminating the temperature oscillations caused by traditional on-off control cycling and from adapting setpoints to actual occupant preferences rather than generic standards.

Response Time

The agent-based system responded to changing conditions far faster than traditional BMS schedules. When a conference room that was expected to be empty filled unexpectedly for an unscheduled meeting, the zone agent detected the occupancy increase through CO2 rise within 90 seconds and began increasing ventilation and cooling within three minutes. Under the previous system, occupants would have experienced discomfort for 15 to 30 minutes before any adjustment occurred.

Implications for Commercial Real Estate

The Cambridge research has implications beyond a single building. Several commercial real estate operators have expressed interest in scaling the framework across their portfolios. The multi-agent architecture is particularly well-suited to portfolio deployment because each building can run its own agent ensemble while a portfolio-level coordination layer optimizes across buildings for utility demand response programs and carbon reduction targets.

The research team has open-sourced the agent communication protocol and is developing a reference implementation compatible with standard BACnet and Modbus building automation interfaces. This means the framework can be deployed on existing HVAC infrastructure without replacing hardware.

Challenges and Limitations

The researchers acknowledge several limitations. The agent-based system requires significantly more sensor data than traditional BMS, including CO2 sensors in each zone, sub-metered energy monitoring, and reliable network connectivity. Installation costs for the sensing infrastructure in the pilot building were approximately 4.50 dollars per square meter, though the energy savings achieved a payback period of under 18 months.

There are also cybersecurity considerations. An agent-based building control system with internet-connected weather and pricing feeds introduces attack surfaces that isolated BMS systems do not have. The Cambridge team implemented network segmentation and anomaly detection to address this concern but notes that building cybersecurity standards need to evolve alongside the technology.

Frequently Asked Questions

Does Office-in-the-Loop require replacing existing HVAC equipment?

No. The framework operates as a control layer on top of existing HVAC infrastructure. It communicates with chillers, air handling units, and variable air volume boxes through standard BACnet and Modbus protocols. The only hardware additions are supplemental sensors for occupancy detection and zone-level environmental monitoring.

How does the system handle occupant disagreements about temperature?

The Zone Comfort Agent maintains individual preference profiles when possible and applies majority-preference logic in shared spaces. In open-plan offices, the agent targets the setpoint that satisfies the greatest number of occupants within the zone. In conference rooms and private offices, it adapts to the specific occupants detected in the space.

Can this approach work in residential buildings?

The Cambridge research focused on commercial office environments, but the multi-agent architecture is adaptable to multi-unit residential buildings. The key difference is that residential applications require stronger privacy protections for occupancy data and individual unit-level comfort control rather than zone-based management.

What is the expected ROI for deploying this system?

Based on the pilot results, buildings with annual HVAC energy costs above 8 dollars per square meter can expect a full ROI within 18 to 24 months. Buildings in regions with time-of-use electricity pricing or demand response programs see faster payback due to the load-shifting capabilities of the energy optimization agent.

---

Source: Cambridge University Engineering Department — Office-in-the-Loop Research, International Energy Agency — Building Energy Efficiency, ASHRAE — Thermal Comfort Standards