Multi-Agent Home Energy Management Assistant

arXiv:2602.15219 · cs.HC · Submitted 2026-02-16 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: We're looking at the paper: "Multi-Agent Home Energy Management Assistant".

Jane: Here's the gist of the paper. Multi-Agent Home Energy Management Assistant presents HEMA, an open-source, multi-agent system designed to enable sustained human-AI collaboration across diverse home energy management tasks through natural language interaction.

Tom: This system combines large language model reasoning with 36 purpose-built domain-specific tools through a three-layer architecture.

Jane: System Architecture and Core Components HEMA implements a three-layer architecture consisting of a Frontend, Backend, and Multi-Agent System.

Tom: The frontend is built with React, Vite, and Tailwind CSS to enable natural language interactions.

Paper summary: Tom: So, wrapping up on "Multi-Agent Home Energy Management Assistant," the core of this work is creating this multi agent system that allows for sustained human-AI collaboration across diverse tasks using a three layer architecture. It’s about giving users a conversational interface for analysis, knowledge retrieval, and device control all at once.

Jane: The authors are showing how to combine large language model reasoning with thirty six purpose built tools to achieve this collaboration, focusing on preserving context over multiple turns during these interactions one <ref:2602.15219#pg2>. It moves beyond single turn queries.

Lu: What they’re really pushing is the idea of specialized agents—Analysis, Knowledge, and Control—that are coordinated by a self-consistency classifier using chain-of-thought reasoning to decide which tool to use next. That’s the architectural innovation here five <ref:2602.15219#pg3>.

Meng: From a deployment standpoint, it's an open source project built on LangChain and LangGraph with FastAPI, which gives the community flexibility in how they integrate real data sources or device APIs.

Lalam: The system is designed to be fully customizable because it’s open source, which means people can tailor those thirty six domain-specific tools exactly to their needs two.

Jane: So what does this mean for the world beyond just home energy? It suggests a future where personalized, sustained engagement with complex energy data becomes much more accessible through natural language.

Tom: It means homeowners can identify high consumption appliances, understand time of use rate implications, explore rebate eligibility, and control their smart devices all through a single conversational interface one <ref:2602.15219#pg2>.

Lu: The implication is that we move from reactive systems to truly proactive ones where the AI can guide users through complex energy management decisions over time.

Jane: It’s a framework for building these kinds of personalized energy advisors, moving past just simple chatbots toward systems that actually understand and act on real home data four.

Conclusion: Tom: So we've seen how this system actually works, and now we gotta talk about what it's called—Multi-Agent Home Energy Management Assistant—and who cooked up this thing and why it matters for us outside of just being a neat piece of software.

Jane: Yeah, the authors are showing how to take a bunch of different AI agents, each with its own job, and make them talk to each other to manage your home energy stuff like analyzing usage, checking rebates, or controlling devices.

Lu: The big idea is combining that reasoning power of the large language model with those specific tools so the AI doesn't just guess; it actually uses specialized knowledge for analysis and control.

Meng: For us on the engineering side, it’s about modularity because you can swap out or add new tools easily without rewriting everything, which is important when you're trying to ship something that has to be reliable at scale.

Lalam: I'm pretty sure the authors are showing that this setup lets a human have a really deep conversation with the system, rather than just clicking buttons in an app.

Tom: It’s about moving from simple commands to sustained collaboration where you can ask complex energy questions and get layered answers from different parts of the system.

Jane: It means for homeowners, it could be like having a personal expert who understands their energy use and can actually help them figure out how to save money or use their devices smarter.

Lu: The potential is using this as a platform where you can experiment with how different agent designs work for different home management tasks.

Tom: It’s an open source research tool that shows the direction the field is headed toward more complex, multi-step AI assistants for real-world tasks.

Jane: It sets up a good starting point for anyone who wants to build something similar, but it's still a proof of concept using household data, so we gotta remember that limitation.

Meng: Exactly. It’s a solid blueprint, but getting it running with actual utility data and real smart meters is the next big hurdle for moving this into the real world.

Lalam: The paper suggests that this multi-agent approach is a way to handle the complexity of home energy management in a very structured and interactive way.

University of Arizona

cs.HC

Submitted: 2026-02-16

Updated: 2026-11-05

Comments: 22 pages, 9 figures. Re-set in the HUBS Lab paper layout (HUBS Lab Paper No. 2026-001); content unchanged

Journal ref: SoftwareX 34 (2026) 102633

DOI: 10.1016/j.softx.2026.102633

Code: https://github.com/humanbuildingsynergy/HEMA

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 91/100

The gist: Multi-Agent Home Energy Management Assistant presents HEMA, an open-source, multi-agent system designed to enable sustained human-AI collaboration across diverse home energy management tasks through…

Key concepts

Multi-Agent System
HEMA uses multiple specialized AI agents—Analysis, Knowledge, and Control—that work together. They coordinate their tasks using a reasoning process to handle complex home energy requests. This modular approach allows the system to perform diverse functions like analyzing data or managing devices.
ReAct Paradigm
Each agent follows the ReAct pattern, which means they reason about what needs to be done and then decide which specific tool to use next. This iterative process helps agents break down complex user queries into actionable steps, leading to more accurate and goal-oriented responses.
Agent Specialization
The system divides labor among three distinct agents: Analysis for consumption data, Knowledge for facts like rebates, and Control for smart devices. This specialization ensures that each agent is highly capable in its specific domain while maintaining overall system coherence.

Terminology

Summary

Multi-Agent Home Energy Management Assistant presents HEMA, an open-source, multi-agent system designed to enable sustained human-AI collaboration across diverse home energy management tasks through natural language interaction. p. 2

Where the paper says it

Furthermore, these studies primarily employed accuracy and/or system efficiency metrics [6–9, 12] that may not capture the full spectrum of human-AI interaction quality and system performance. Goal and Significance This paper introduces Home Energy Management Assistant (HEMA), the first open-source, fully customizable multi-agent system for HEM, designed to enable sustained human-AI collaboration across diverse HEM tasks.

This system combines large language model reasoning with 36 purpose-built domain-specific tools through a three-layer architecture. p. 0
Where the paper says it

This paper presents Home Energy Management Assistant (HEMA), the first open-source, multi-agent system enabling sustained human-AI collaboration – multi-turn conversational interactions with preserved context – across diverse home energy management (HEM) tasks – from energy analysis and educational support to smart device control. HEMA combines large language model (LLM) reasoning capabilities with 36 purpose-built domain-specific tools through a three-layer architecture: a web-based conversational interface, a backend API server, and a multi-agent system.

System Architecture and Core Components

HEMA implements a three-layer architecture consisting of a Frontend, Backend, and Multi-Agent System. The frontend is built with React, Vite, and Tailwind CSS to enable natural language interactions. p. 3

Where the paper says it

The frontend provides a user-friendly web interface that enables natural language interactions through a conversational chat interface. It is built with React (JavaScript library that enables building interactive UIs [13]), Vite (a fast build tool [14]), and Tailwind CSS (responsive styling that adapts to different screen sizes [15]).

The backend uses FastAPI to manage requests and supports multiple LLM providers through a provider-agnostic interface. p. 5
Where the paper says it

It serves as the core processing unit that handles incoming requests from the frontend, manages interactions with the multi-agent system, and returns responses. The backend is designed to be flexible and scalable, supporting multiple LLM providers (OpenAI, Ollama, Google, Anthropic) through a provider-agnostic interface with intelligent cascading fallback – if the primary provider fails, the system automatically attempts the next provider in the cascade, ensuring resilience across network failures and API unavailability.

The multi-agent system is the core, implemented using LangChain and LangGraph for modular, stateful agent workflows. p. 5
Where the paper says it

The multi-agent system is the core of HEMA, implemented using LangChain (a framework for building applications with LLMs [17]) and LangGraph (a framework for building agentic applications with state management [18]). These two complementary frameworks enable the construction of modular, extensive AI systems that offer an explicit state graph architecture for defining agent workflows, where transitions are predictable directed edges rather than implicit LLM decisions.

Multi-Agent Specialization and Coordination

The system features three specialized agents: Analysis (energy consumption patterns and cost optimization), Knowledge (educational queries and rebate information), and Control (smart device management and scheduling). These agents are coordinated through a self-consistency classifier that routes user queries using chain-of-thought reasoning. p. 0

Where the paper says it

HEMA combines large language model (LLM) reasoning capabilities with 36 purpose-built domain-specific tools through a three-layer architecture: a web-based conversational interface, a backend API server, and a multi-agent system. The system features three specialized agents – Analysis (energy consumption patterns and cost optimization), Knowledge (educational queries and rebate information), and Control (smart device management and scheduling) – coordinated through a self-consistency classifier that routes user queries using chain-of-thought reasoning.

Each agent follows the ReAct paradigm, reasoning about actions and invoking purpose-built tools based on the query. p. 5
Where the paper says it

Second, once routed, the designated agent – Analysis, Knowledge, Control – processes the query using its specialized tools and reasoning capabilities (their system prompts are available in Appendix B). Each agent follows the ReAct paradigm [21]: (1) reason about required actions based on query and context, (2) invoke purpose-built tools with appropriate 6

Agent Capabilities and Toolsets

The Analysis Agent processes energy consumption data through phases like data loading, interactive tool selection, and response generation with cost estimates. The Knowledge Agent uses Retrieval-Augmented Generation (RAG) over indexed documents to provide grounded, citation-aware responses regarding energy concepts and rebates. p. 6

Where the paper says it

Department of Energy (DOE) guides [22], ENERGY STAR specifications [23], rebate databases; weather APIs from Open-Meteo [24]; built-in knowledge, (2) information retrieval and grounding. The RAG system enables semantic search over energy documents to provide grounded, citation-aware responses.

The Control Agent handles smart device queries via a discovery-first operation, including device validation and safe command execution. p. 6
Where the paper says it

A total of 8 purpose-built tools are available, including document search, energy knowledge retrieval, weather integration, and user context personalization. Control Agent: Handles smart device queries through discovery-first operation: (1) home context gathering (device inventory, capabilities, constraints), (2) device validation and parameter verification (respecting operational limits: thermostat 60–85◦F, mode appropriateness by season), (3) safe command execution with explicit confirmation.

Evaluation Framework and Results

HEMA includes a comprehensive evaluation framework using an LLM-as-simulated-user methodology with 23 objective metrics across task performance, factual accuracy, interaction quality, and system efficiency. The evaluation involved 105 total runs across three user personas and seven realistic scenarios. p. 8

Where the paper says it

In total, three user personas and seven realistic scenarios are included (details in Appendix C), covering sample and representative HEM tasks. A total of 23 objective metrics are defined across four categories: task performance, factual accuracy, interaction quality, and system efficiency (see Appendix D for details).

Overall evaluation outcomes showed a 98. p. 8
Where the paper says it

In the end, this comprehensive evaluation framework enables a rigorous assessment of the multifaceted nature of human-AI collaboration in the context of HEMS. Illustrative Examples This section demonstrates HEMA through three realistic conversations from actual system evaluations, and a summary of overall evaluation outcomes using the LLM-as-simulated-user methodology.

1% goal achievement rate and 95. p. 2
Where the paper says it

Furthermore, these studies primarily employed accuracy and/or system efficiency metrics [6–9, 12] that may not capture the full spectrum of human-AI interaction quality and system performance. Goal and Significance This paper introduces Home Energy Management Assistant (HEMA), the first open-source, fully customizable multi-agent system for HEM, designed to enable sustained human-AI collaboration across diverse HEM tasks.

9% factual accuracy. p. 0
Where the paper says it

This architecture enables various energy analyses, adaptive explanations, and streamlined device control. HEMA also includes a comprehensive evaluation framework using an LLM-as-simulated-user methodology with 23 objective metrics across task performance, factual accuracy, interaction quality, and system efficiency, allowing systematic testing across diverse scenarios and user personas without requiring extensive human subject testing.

Impact and Future Directions

HEMA supports homeowners by enabling engagement with energy data through natural language, allowing them to identify high-consumption appliances, understand time-of-use rate implications, explore rebate eligibility, and control smart devices. The system benefits utility companies by supporting customer-side demand flexibility through device control capabilities. p. 6

Where the paper says it

Control Agent: Handles smart device queries through discovery-first operation: (1) home context gathering (device inventory, capabilities, constraints), (2) device validation and parameter verification (respecting operational limits: thermostat 60–85◦F, mode appropriateness by season), (3) safe command execution with explicit confirmation. Consequently, this agent enables scheduling (shift loads to off-peak hours), automation (demand response integration), and real-time monitoring with energy consumption tracking across the device fleet.

Future work includes expanding the tool ecosystem, enhancing query routing with more advanced classification techniques, and integrating real-time data sources like smart meters for dynamic control. The paper emphasizes that HEMA serves as an extensible research platform for the HEM community. p. 0
Where the paper says it

HEMA also includes a comprehensive evaluation framework using an LLM-as-simulated-user methodology with 23 objective metrics across task performance, factual accuracy, interaction quality, and system efficiency, allowing systematic testing across diverse scenarios and user personas without requiring extensive human subject testing. Through demonstrations using real-world household energy consumption data, we show how HEMA supports informed decision-making and active engagement in HEM, highlighting its potential as a user-friendly, adaptable tool for residential deployment and as a research platform for HEM innovation.

Limitations and Safety Considerations

HEMA inherits general LLM limitations, including non-deterministic outputs and potential hallucination [31]. While the multi-agent architecture mitigates some risks, safety guardrails like user confirmation for device control actions do not fully eliminate LLM-related risks for safety-critical operations. p. 0

Where the paper says it

HEMA combines large language model (LLM) reasoning capabilities with 36 purpose-built domain-specific tools through a three-layer architecture: a web-based conversational interface, a backend API server, and a multi-agent system. The system features three specialized agents – Analysis (energy consumption patterns and cost optimization), Knowledge (educational queries and rebate information), and Control (smart device management and scheduling) – coordinated through a self-consistency classifier that routes user queries using chain-of-thought reasoning.

The current version is configured as a single-household proof-of-concept using static data, and real-world deployment would require integration with real-time data sources and utility APIs. p. 0
Where the paper says it

HEMA also includes a comprehensive evaluation framework using an LLM-as-simulated-user methodology with 23 objective metrics across task performance, factual accuracy, interaction quality, and system efficiency, allowing systematic testing across diverse scenarios and user personas without requiring extensive human subject testing. Through demonstrations using real-world household energy consumption data, we show how HEMA supports informed decision-making and active engagement in HEM, highlighting its potential as a user-friendly, adaptable tool for residential deployment and as a research platform for HEM innovation.

Conclusion

HEMA provides both a deployable residential tool and a research platform for advancing LLM-integrated HEMS by combining LLM reasoning with domain-specific tools to support diverse home energy management tasks. The system demonstrates how informed decision-making and sustained engagement can be supported in residential settings through practical demonstrations using real-world household energy consumption data. p. 0

Where the paper says it

HEMA also includes a comprehensive evaluation framework using an LLM-as-simulated-user methodology with 23 objective metrics across task performance, factual accuracy, interaction quality, and system efficiency, allowing systematic testing across diverse scenarios and user personas without requiring extensive human subject testing. Through demonstrations using real-world household energy consumption data, we show how HEMA supports informed decision-making and active engagement in HEM, highlighting its potential as a user-friendly, adaptable tool for residential deployment and as a research platform for HEM innovation.

References

[1] X. Jin, K. Baker, D. Christensen, and S. Isley. Foresee: A user-centric home energy management system for human-building interaction. p. 2

Where the paper says it

Furthermore, these studies primarily employed accuracy and/or system efficiency metrics [6–9, 12] that may not capture the full spectrum of human-AI interaction quality and system performance. Goal and Significance This paper introduces Home Energy Management Assistant (HEMA), the first open-source, fully customizable multi-agent system for HEM, designed to enable sustained human-AI collaboration across diverse HEM tasks.

[2] B. Zhou, W. Li, K.W. Chan, Y. Cao, Y Kuang, X Liu, and X Wang. Smart home energy management systems: Concept, configurations, and scheduling strategies. p. 0

Where the paper says it

Multi-Agent Home Energy Management Assistant Wooyoung Junga,∗ aCivil and Architectural Engineering and Mechanics, University of Arizona, Tucson, Arizona, USA Abstract Existing home energy management systems conceptualize occupants as passive recipients of energy information and control, which limits their ability to effectively support informed decision-making and sustained engagement. This paper presents Home Energy Management Assistant (HEMA), the first open-source, multi-agent system enabling sustained human-AI collaboration – multi-turn conversational interactions with preserved context – across diverse home energy management (HEM) tasks – from energy analysis and educational support to smart device control.

[3] M A Hannan et al., A review of internet of energy based building energy management systems: Issues and recommendations. p. 1

Where the paper says it

Background and Related Work Homeowners increasingly face complex energy management decisions driven by distributed energy resources (rooftop solar, home battery), dynamic utility rate structures (time-of-use (TOU) pricing), and high-power end-use technologies (electric vehicle (EV) charging) [1]. Existing home energy management systems (HEMS) rely on centralized optimization and static rule-based logic, offering limited user interaction and adaptability.

[4] M Stogia et al., A scalable and user-friendly framework integrating IoT and digital twins for home energy management systems. p. 0

Where the paper says it

Through demonstrations using real-world household energy consumption data, we show how HEMA supports informed decision-making and active engagement in HEM, highlighting its potential as a user-friendly, adaptable tool for residential deployment and as a research platform for HEM innovation. Keywords: Home Energy Management, Multi-Agent System, Human-AI ∗Corresponding author Email address: wooyoung@arizona.

[5] W Jung. p. 0

Where the paper says it

Keywords: Home Energy Management, Multi-Agent System, Human-AI ∗Corresponding author Email address: wooyoung@arizona. edu (Wooyoung Jung) arXiv:2602.

Chain-of-thought prompting for human-centered home energy management. p. 7
Where the paper says it

<Persona & Task> You are classifying a user query for Home Energy Management Assistant. Determine the AGENT of this query using chain-of-thought reasoning <Chain-of-Thought Instructions> <Response Format> Apply the reasoning process above, then provide your answer as JSON: json { "result": "Core request: [what user wants].

[6] T He and F Jazizadeh. Context-aware LLM-based AI agents for human-centered energy management systems in smart buildings. p. 0

Where the paper says it

This paper presents Home Energy Management Assistant (HEMA), the first open-source, multi-agent system enabling sustained human-AI collaboration – multi-turn conversational interactions with preserved context – across diverse home energy management (HEM) tasks – from energy analysis and educational support to smart device control. HEMA combines large language model (LLM) reasoning capabilities with 36 purpose-built domain-specific tools through a three-layer architecture: a web-based conversational interface, a backend API server, and a multi-agent system.

[7] J Rey-Jouanchicot et al., Leveraging large language models for enhanced personalised user experience in smart homes. p. 1

Where the paper says it

Under this growing complexity, there is a critical need for intelligent, user-centered tools that can support households in understanding and analyzing their energy use, improving their energy literacy, and empowering them to make informed decisions [4]. The transformative advancements of large language models (LLMs) in recent years have opened new possibilities for developing interactive HEMS, 2

[8] R E Makroum et al., Agentic AI home energy management system: A large language model framework for residential load scheduling. p. 0

Where the paper says it

HEMA combines large language model (LLM) reasoning capabilities with 36 purpose-built domain-specific tools through a three-layer architecture: a web-based conversational interface, a backend API server, and a multi-agent system. The system features three specialized agents – Analysis (energy consumption patterns and cost optimization), Knowledge (educational queries and rebate information), and Control (smart device management and scheduling) – coordinated through a self-consistency classifier that routes user queries using chain-of-thought reasoning.

[9] F Michelon et al., Large language model interface for home energy management systems. p. 0

Where the paper says it

This paper presents Home Energy Management Assistant (HEMA), the first open-source, multi-agent system enabling sustained human-AI collaboration – multi-turn conversational interactions with preserved context – across diverse home energy management (HEM) tasks – from energy analysis and educational support to smart device control. HEMA combines large language model (LLM) reasoning capabilities with 36 purpose-built domain-specific tools through a three-layer architecture: a web-based conversational interface, a backend API server, and a multi-agent system.

[10] A Papaioannou et al., GUIDE: A prescriptive hybrid AI framework for energy-efficient appliances usage through behavioral modeling and LLM guidance. p. 9

Where the paper says it

Example 2: Confused Newcomer – Rebate Inquiry with Educational Support. This example demonstrates how HEMA provides educational and information support and personalized guidance for a user who is new to energy concepts and interested in exploring rebates for energy-efficient appliances.

[11] N V Gkalinikis et al., RHEA: Residential home energy advisor. p. 0

Where the paper says it

Through demonstrations using real-world household energy consumption data, we show how HEMA supports informed decision-making and active engagement in HEM, highlighting its potential as a user-friendly, adaptable tool for residential deployment and as a research platform for HEM innovation. Keywords: Home Energy Management, Multi-Agent System, Human-AI ∗Corresponding author Email address: wooyoung@arizona.

[12] T He and F Jazizadeh. LLM-based building energy management assistants. p. 0

Where the paper says it

This paper presents Home Energy Management Assistant (HEMA), the first open-source, multi-agent system enabling sustained human-AI collaboration – multi-turn conversational interactions with preserved context – across diverse home energy management (HEM) tasks – from energy analysis and educational support to smart device control. HEMA combines large language model (LLM) reasoning capabilities with 36 purpose-built domain-specific tools through a three-layer architecture: a web-based conversational interface, a backend API server, and a multi-agent system.

[13] Inc Meta Platforms. p. 2

Where the paper says it

However, existing LLMintegrated HEMS are limited in their system design and evaluation. Specifically, they lack specialized agent architectures [9] despite the diversity of HEM tasks, they rely on pre-built platforms that constrain developer customization [6, 12], and evaluate single-turn query-response trials that may not reflect actual human-AI interactions [6–8, 12].

The React Framework for the Web. p. 3
Where the paper says it

The architecture follows a modular design enabling independent development, testing, and deployment of each component. Frontend Backend Multi-Agent AI Framework (LangChain + LangGraph) User Interface (React) Web Framework (FastAPI) Classifier Clarification Agent-based Response & Action Generation Fallback Handler Analysis Agent Knowledge Agent Control Agent Response & Action Smart  Home API* Not implemented *Application Programming Interface Database Figure 1: HEMA three-layer architecture showing frontend, backend, and multi-agent system.

[14] Evan You. p. 7

Where the paper says it

<Persona & Task> You are classifying a user query for Home Energy Management Assistant. Determine the AGENT of this query using chain-of-thought reasoning <Chain-of-Thought Instructions> <Response Format> Apply the reasoning process above, then provide your answer as JSON: json { "result": "Core request: [what user wants].

Vite: Next generation frontend tooling. p. 5
Where the paper says it

It serves as the core processing unit that handles incoming requests from the frontend, manages interactions with the multi-agent system, and returns responses. The backend is designed to be flexible and scalable, supporting multiple LLM providers (OpenAI, Ollama, Google, Anthropic) through a provider-agnostic interface with intelligent cascading fallback – if the primary provider fails, the system automatically attempts the next provider in the cascade, ensuring resilience across network failures and API unavailability.

[15] Tailwind Labs. p. 1

Where the paper says it

com/ humanbuildingsynergy/HEMA C4 Legal Code License GNU General Public License v3. 0) C5 Code versioning system used Git C6 Software code languages, tools, and services used LangChain, LangGraph, FastAPI, React, Vite, Tailwind CSS C7 Compilation requirements, operating environments & dependencies Python 3.

Tailwind CSS. p. 1
Where the paper says it

com/ humanbuildingsynergy/HEMA C4 Legal Code License GNU General Public License v3. 0) C5 Code versioning system used Git C6 Software code languages, tools, and services used LangChain, LangGraph, FastAPI, React, Vite, Tailwind CSS C7 Compilation requirements, operating environments & dependencies Python 3.

[16] Sebastián Ramírez. FastAPI.

[17] LangChain Foundation. p. 3

Where the paper says it

The primary contributions of this work are: (1) providing a foundation for developing customizable multi-agent systems for HEM, (2) offering a fully open-source system ready for community use, extension, and residential deployment, and (3) facilitating the evaluation of LLM-integrated HEMS through a comprehensive testing framework with diverse metrics. Software Description 2.

LangChain.

[18] J Wang and Z Duan. Agent AI with LangGraph: A modular framework for enhancing machine translation using large language models. p. 5

Where the paper says it

The multi-agent system is the core of HEMA, implemented using LangChain (a framework for building applications with LLMs [17]) and LangGraph (a framework for building agentic applications with state management [18]). These two complementary frameworks enable the construction of modular, extensive AI systems that offer an explicit state graph architecture for defining agent workflows, where transitions are predictable directed edges rather than implicit LLM decisions.

[19] X Wang et al., Self-consistency improves chain of thought reasoning in language models. p. 0

Where the paper says it

HEMA combines large language model (LLM) reasoning capabilities with 36 purpose-built domain-specific tools through a three-layer architecture: a web-based conversational interface, a backend API server, and a multi-agent system. The system features three specialized agents – Analysis (energy consumption patterns and cost optimization), Knowledge (educational queries and rebate information), and Control (smart device management and scheduling) – coordinated through a self-consistency classifier that routes user queries using chain-of-thought reasoning.

[20] J Wei et al., Emergent abilities of large language models. p. 1

Where the paper says it

Under this growing complexity, there is a critical need for intelligent, user-centered tools that can support households in understanding and analyzing their energy use, improving their energy literacy, and empowering them to make informed decisions [4]. The transformative advancements of large language models (LLMs) in recent years have opened new possibilities for developing interactive HEMS, 2

[21] S Yao et al., React: Synergizing reasoning and acting in language models. p. 2

Where the paper says it

HEMA offers the following key features: • Conversational Interface with Flexible Backend: A web-based conversational interface enabling natural language interactions with support for multi-turn conversations, backed by a provider-agnostic architecture supporting multiple LLM providers and deployable across cloud platforms or local machines without vendor lock-in. • Multi-Agent Specialization with Intelligent Routing: Three domain-specific agents (Analysis for energy patterns and optimization, Knowledge for education and rebates, Control for device management), each equipped with purpose-built tools following the reasoning and action (ReAct) paradigm, coordinated through a self-consistency classifier using chain-of-thought (CoT) reasoning with majority voting.

[22] U S Department of Energy. p. 6

Where the paper says it

Knowledge Agent: Responds to general information and educational queries using a two-phase approach: (1) semantic source selection via retrievalaugmented generation (RAG) over indexed documents including the U. Department of Energy (DOE) guides [22], ENERGY STAR specifications [23], rebate databases; weather APIs from Open-Meteo [24]; built-in knowledge, (2) information retrieval and grounding.

U s department of energy: Energy efficiency and renewable energy. p. 6
Where the paper says it

Knowledge Agent: Responds to general information and educational queries using a two-phase approach: (1) semantic source selection via retrievalaugmented generation (RAG) over indexed documents including the U. Department of Energy (DOE) guides [22], ENERGY STAR specifications [23], rebate databases; weather APIs from Open-Meteo [24]; built-in knowledge, (2) information retrieval and grounding.

[23] U S Environmental Protection Agency. ENERGY STAR: Trusted partnership for a cleaner environment. p. 6

Where the paper says it

Knowledge Agent: Responds to general information and educational queries using a two-phase approach: (1) semantic source selection via retrievalaugmented generation (RAG) over indexed documents including the U. Department of Energy (DOE) guides [22], ENERGY STAR specifications [23], rebate databases; weather APIs from Open-Meteo [24]; built-in knowledge, (2) information retrieval and grounding.

[24] Open-Meteo. p. 6

Where the paper says it

Knowledge Agent: Responds to general information and educational queries using a two-phase approach: (1) semantic source selection via retrievalaugmented generation (RAG) over indexed documents including the U. Department of Energy (DOE) guides [22], ENERGY STAR specifications [23], rebate databases; weather APIs from Open-Meteo [24]; built-in knowledge, (2) information retrieval and grounding.

Open-Meteo: Free weather API. p. 6
Where the paper says it

Knowledge Agent: Responds to general information and educational queries using a two-phase approach: (1) semantic source selection via retrievalaugmented generation (RAG) over indexed documents including the U. Department of Energy (DOE) guides [22], ENERGY STAR specifications [23], rebate databases; weather APIs from Open-Meteo [24]; built-in knowledge, (2) information retrieval and grounding.

[30] Samsung. Let’s build a connected future. p. 3

Where the paper says it

The frontend provides a user-friendly web interface that enables natural language interactions through a conversational chat interface. It is built with React (JavaScript library that enables building interactive UIs [13]), Vite (a fast build tool [14]), and Tailwind CSS (responsive styling that adapts to different screen sizes [15]).

[31] L Huang et al., A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. p. 1

Where the paper says it

Under this growing complexity, there is a critical need for intelligent, user-centered tools that can support households in understanding and analyzing their energy use, improving their energy literacy, and empowering them to make informed decisions [4]. The transformative advancements of large language models (LLMs) in recent years have opened new possibilities for developing interactive HEMS, 2

Appendix A. p. 5

Where the paper says it

When ambiguity arises (ties in voting), the system presents clarification options rather than misrouting, preventing cascade failures in downstream agents. Second, once routed, the designated agent – Analysis, Knowledge, Control – processes the query using its specialized tools and reasoning capabilities (their system prompts are available in Appendix B).

p. 5
Where the paper says it

When ambiguity arises (ties in voting), the system presents clarification options rather than misrouting, preventing cascade failures in downstream agents. Second, once routed, the designated agent – Analysis, Knowledge, Control – processes the query using its specialized tools and reasoning capabilities (their system prompts are available in Appendix B).

1 Analysis Agent Tools

The Analysis agent includes 18 tools organized into four categories such as Data loading tools, Consumption and period analysis tools, Pattern and trend analysis tools, and Rate, solar, and summary tools. These include a tool named load energy data for ingesting appliance-level energy consumption data. p. 6

Where the paper says it

Their detailed capabilities are as follows: Analysis Agent: Processes energy consumption data through three phases: (1) data loading and metadata retrieval (energy CSV files, utility rates, appliance thresholds), (2) interactive tool selection (consumption statistics, appliance ranking, peak hour analysis, usage trends), (3) response generation with cost estimates and actionable recommendations. Enables multi-step reasoning: load data → identify consumers → analyze rates → calculate savings → recommend actions with quantified impact.

Appendix A.3 Control Agent Tools

The Control agent includes 10 tools for device management and automation, such as Device discovery and status tools, Control actions tools, Scheduling and automation tools. These include a tool named control device for executing control commands like adjusting setpoints. p. 0

Where the paper says it

Multi-Agent Home Energy Management Assistant Wooyoung Junga,∗ aCivil and Architectural Engineering and Mechanics, University of Arizona, Tucson, Arizona, USA Abstract Existing home energy management systems conceptualize occupants as passive recipients of energy information and control, which limits their ability to effectively support informed decision-making and sustained engagement. This paper presents Home Energy Management Assistant (HEMA), the first open-source, multi-agent system enabling sustained human-AI collaboration – multi-turn conversational interactions with preserved context – across diverse home energy management (HEM) tasks – from energy analysis and educational support to smart device control.

Appendix B System Prompts

Appendix B provides excerpts of the system prompts that define each agent’s role, capabilities, workflow, and constraints. These prompts are central to HEMA’s behavior as they instruct the LLM on how to reason and which tools to invoke. p. 5

Where the paper says it

Second, once routed, the designated agent – Analysis, Knowledge, Control – processes the query using its specialized tools and reasoning capabilities (their system prompts are available in Appendix B). Each agent follows the ReAct paradigm [21]: (1) reason about required actions based on query and context, (2) invoke purpose-built tools with appropriate 6

Appendix C Evaluation Methodology

The evaluation methodology employs an LLM-as-simulated-user approach, covering three user personas and seven realistic scenarios. This approach allows for systematic testing without requiring extensive human subject recruitment. p. 0

Where the paper says it

This architecture enables various energy analyses, adaptive explanations, and streamlined device control. HEMA also includes a comprehensive evaluation framework using an LLM-as-simulated-user methodology with 23 objective metrics across task performance, factual accuracy, interaction quality, and system efficiency, allowing systematic testing across diverse scenarios and user personas without requiring extensive human subject testing.

The framework defines 23 objective metrics across four categories to rigorously assess the system's performance.

Appendix E Detailed Evaluation Results

Detailed results from 105 test runs show that all scenarios achieved 100% goal achievement except vacation preparation (86.7%). The high variance in turns to completion reflects the range of scenario complexity, with multi-device orchestration requiring longer processing times. Performance was highest for knowledge-oriented scenarios (rebate inquiry and utility rate: 100%). p. 9

Where the paper says it

The last part of Response #1 was cut off for brevity. Example 2: Confused Newcomer – Rebate Inquiry with Educational Support.

Appendix E. p. 5

Where the paper says it

When ambiguity arises (ties in voting), the system presents clarification options rather than misrouting, preventing cascade failures in downstream agents. Second, once routed, the designated agent – Analysis, Knowledge, Control – processes the query using its specialized tools and reasoning capabilities (their system prompts are available in Appendix B).

3 Interaction Quality and System Efficiency

The actionable recommendation ratio is 64.1%, reflecting HEMA’s design to balance specific, data-backed recommendations with educational context. p. 2

Where the paper says it

Here, sustained collaboration refers to multi-turn interactions where conversational context is preserved, enabling users to iteratively explore energy questions and act on recommendations. HEMA offers the following key features: • Conversational Interface with Flexible Backend: A web-based conversational interface enabling natural language interactions with support for multi-turn conversations, backed by a provider-agnostic architecture supporting multiple LLM providers and deployable across cloud platforms or local machines without vendor lock-in.

The average response latency is 35. p. 9
Where the paper says it

31 kWh (55. 0% of total), Daily Average: 76.

6 s ± 103. p. 9
Where the paper says it

1 kWh, Estimated Monthly: 1,413 kWh. • Electric Vehicle Charger: Total Consumption: 103.

9 s, and the tool calls per run average 18. p. 9
Where the paper says it

31 kWh (55. 0% of total), Daily Average: 76.

4 ± 36.5.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Adapt query routing for greater precision by developing more advanced classification techniques beyond current chain-of-thought reasoning, potentially to reduce misrouting in ambiguous queries as mentioned in Section 2.2 of the paper.

  2. Enhance response quality for technical users by increasing the Jargon Explanation Rate through fine-tuning agent system prompts based on scenario analysis, ensuring that explanations are more tailored to specific user expertise levels and goals.

  3. Improve robustness against LLM risks for safety-critical operations by implementing stricter validation protocols within the Control Agent, specifically addressing safety-critical device control operations as noted in Section 5 of the Limitations.

  4. Increase system efficiency for complex tasks by developing multi-agent coordination strategies to better handle complex multi-step tasks, which currently show high latency during scenarios like vacation preparation (Page 28).

  5. Expand real-world applicability by integrating Model Context Protocols (MCP) for real-time data streaming to replace static appliance data, addressing the limitation that HEMA is currently configured as a single-household proof-of-concept.

Abstract

Existing home energy management systems conceptualize occupants as passive recipients of energy information and control, which limits their ability to effectively support informed decision-making and sustained engagement. This paper presents Home Energy Management Assistant (HEMA), the first open-source, multi-agent system enabling sustained human-AI collaboration - multi-turn conversational interactions with preserved context - across diverse home energy management (HEM) tasks - from energy analysis and educational support to smart device control. HEMA combines large language model (LLM) reasoning capabilities with 36 purpose-built domain-specific tools through a three-layer architecture: a web-based conversational interface, a backend API server, and a multi-agent system. The system features three specialized agents - Analysis (energy consumption patterns and cost optimization), Knowledge (educational queries and rebate information), and Control (smart device management and scheduling) - coordinated through a self-consistency classifier that routes user queries using chain-of-thought reasoning. This architecture enables various energy analyses, adaptive explanations, and streamlined device control. HEMA also includes a comprehensive evaluation framework using an LLM-as-simulated-user methodology with 23 objective metrics across task performance, factual accuracy, interaction quality, and system efficiency, allowing systematic testing across diverse scenarios and user personas without requiring extensive human subject testing. Through demonstrations using real-world household energy consumption data, we show how HEMA supports informed decision-making and active engagement in HEM, highlighting its potential as a user-friendly, adaptable tool for residential deployment and as a research platform for HEM innovation.

Sources

Related papers