The AGENT project, funded by the German federal government through the Federal Ministry of Economic Affairs and Climate Action and managed by the Jülich Project Office, ran from 1 May 2019 to 31 December 2022. Its goal was to improve user comfort and cut CO₂ emissions in non‑residential buildings by replacing conventional, rule‑based control with a distributed, agent‑based system that can model and coordinate the complex interactions of modern building technologies. The research was carried out by a consortium of RWTH Aachen University, Friedrich‑Alexander‑University Erlangen‑Nürnberg, and Robert Bosch GmbH. RWTH provided the core software development, simulation, and model‑library work; FAU supplied advanced control theory, optimisation algorithms, and comparative studies; Bosch implemented the demonstrator in a real building, performed monitoring, and evaluated field performance.

Technically, the project produced a complete Python framework that allows developers to create, test, and deploy building agents. The core library, AgentLib, defines a standard agent interface and supports communication over multiple protocols, including Azure IoT Hub, Fiware, and a custom CloneMAP protocol. Together with the CloneMAP library, the framework enables plug‑and‑play integration of new agents without requiring knowledge of the entire system. The project also delivered BESMod, a modular building‑energy‑model library that incorporates a CO₂ model into the reduced‑order AixLib framework, allowing simultaneous optimisation of energy consumption and emissions.

A key contribution was the development of a software‑in‑the‑loop (SIL) test bench that simulates building dynamics, agent interactions, and network communication. This virtual test environment was used to evaluate and compare several distributed model‑predictive‑control (MPC) strategies. Hierarchical MPC, the alternating‑direction‑method‑of‑multipliers (ADMM) algorithm, sensitivity‑based MPC, and negotiation‑based MPC were implemented in AgentLib and benchmarked against each other. The comparison, performed over a seven‑day closed‑loop simulation, demonstrated that the ADMM‑based approach achieved the lowest energy consumption while maintaining comfort constraints, and that sensitivity‑based MPC offered a favourable trade‑off between computational load and performance.

The demonstrator, installed in a Bosch research building in Renningen, integrated the full stack: automatically generated semantic building models, Azure cloud connectivity, and the plug‑and‑play service. Real‑time monitoring of temperature, humidity, CO₂, and power consumption confirmed the simulation results: the agent‑based system reduced peak power demand by up to 15 % and lowered CO₂ emissions by roughly 10 % compared with a conventional thermostat strategy, while keeping occupant comfort within the prescribed limits. The system’s scalability was validated by adding new agents for additional subsystems without re‑engineering the existing code base.

Overall, the AGENT project delivered a reusable, open‑source agent framework, a comprehensive building‑model library, and a validated distributed MPC strategy that together enable intelligent, robust control of complex energy systems in non‑residential buildings. The collaboration between academia and industry ensured that the solutions are both scientifically rigorous and practically deployable, paving the way for wider adoption of agent‑based building management in the future.