Pith. sign in

REVIEW 3 major objections 45 references

Commercial language-model agents with native tool calling can schedule multi-appliance home loads near the mathematical optimum and capture almost all available savings from dynamic tariffs.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-11 17:06 UTC pith:5VOW2UXS

load-bearing objection Solid agentic-HEMS systems paper: function-calling commercial models hit near-MILP multi-appliance schedules on Agile, and the constraint suite is the real contribution; the £1,270/year number is a soft projection, not the load-bearing result. the 3 major comments →

arxiv 2607.04569 v1 pith:5VOW2UXS submitted 2026-07-06 eess.SY cs.SY

LLMs for Agentic Home Energy Management

Sokipriala Jonah This is my paper
classification eess.SY cs.SY
keywords agentic AIhome energy management systemslarge language modelsdemand responseload schedulingsolar self-consumptiondynamic retail tariffsLLM benchmarking
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Home energy management can cut bills and support demand response, but people rarely adopt systems that force them to turn everyday preferences into technical constraints. This paper tests whether large language model agents can close that gap by taking natural-language requests and committing schedules for washing machines, dishwashers, and EV chargers against live half-hourly UK retail prices, weather, and solar forecasts. A single tool-calling agent is scored against a mixed-integer linear program that co-optimizes net cost with solar self-consumption and household power limits. With native function calling, three commercial models achieve perfect scheduling success and near-optimal cost; free-form text actions sharply reduce reliability, and constraint-conflict tests show the cheapest model is not the safest. Over a week of rolling deployment the agents capture 96.7–98.0% of the solver’s savings versus an off-peak timer, projecting about £1,270 a year for a representative household.

Core claim

With native function calling, GPT-4o-mini, Gemini 2.5 Flash, and Claude Sonnet 4.6 all achieve 100% multi-appliance scheduling success and near-MILP optimality across twelve Agile tariff days; over a week-long rolling deployment they capture 96.7–98.0% of extended-MILP oracle savings, while cost-optimal and safety-optimal models diverge under power-cap, deadline, and infeasibility conflicts.

What carries the argument

A single tool-calling ReAct agent that retrieves live prices, weather, and PV forecasts, then commits schedules only through structured function calls, evaluated against an extended mixed-integer linear program that couples appliances via shared solar and a household power cap.

Load-bearing premise

The roughly £1,270 annual savings figure treats one archived evaluation week as representative of a full year of prices, weather, and household use.

What would settle it

Run the same agents over a multi-season price and weather archive, or a real household pilot, and check whether they still capture roughly 97% of the extended-MILP savings versus an off-peak timer; a large drop, especially under atypical volatility or solar regimes, would falsify the headline savings claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Native function calling can make a flat single agent sufficient for small residential multi-appliance coordination without hierarchical specialist agents.
  • Every LLM-committed schedule should pass a model-independent feasibility validator before device actuation.
  • Weather-aware co-optimization should be regime-dependent: useful on overcast days, sometimes harmful when import prices go negative.
  • Model choice for deployment should trade per-event cost against constraint-compliance under conflict, not only average optimality.
  • Households on dynamic retail tariffs with rooftop PV can approach solver-level savings without encoding technical parameters themselves.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Many earlier open-source agentic HEMS failures may be action-interface artifacts rather than pure reasoning limits, given how sharply text-parsed modes degrade here.
  • The propose-then-validate pattern (LLM suggests, deterministic checker gates) likely generalizes to other safety-critical home automation beyond energy.
  • Scaling the flat agent to thermal loads, batteries, and longer horizons will reveal when hierarchy becomes necessary again.
  • If multi-season field data confirm the savings share, conversational HEMS become a high-ROI consumer product rather than a research demo.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The paper proposes a single tool-calling ReAct agent for multi-appliance home energy management that schedules washing machine, dishwasher, and EV charging against live Octopus Agile half-hourly prices, weather forecasts, and a PV generation model, with an extended MILP as ground truth. It benchmarks GPT-4o-mini, Gemini 2.5 Flash, and Claude Sonnet 4.6 across 12 tariff days, a constraint-conflict suite (S1–S6), weather-aware co-optimization, and a week-long rolling deployment. The central empirical claims are that native function calling yields 100% multi-appliance success and near-MILP optimality, that text-parsed action interfaces sharply reduce reliability (including Claude’s conversational short-circuiting), that models split on safety-critical constraint conflicts, that weather co-optimization is regime-dependent, and that agents capture 96.7–98.0% of oracle savings over one simulated week, from which the authors project ~£1,270 annual savings versus an off-peak timer.

Significance. If the controlled results hold, the paper is a useful contribution to agentic HEMS: it shows that commercial function-calling LLMs can close multi-appliance coordination gaps reported for open-source systems, that action-interface design is a first-order reliability factor, and that constraint-conflict behaviour—not benign-case optimality—should drive model selection. Strengths include an explicit linearized MILP ground truth (Appendix A), archived price/weather inputs, temperature-0 repeated runs, statistical tests in Experiment 1, a reusable adversarial scenario suite, transparent failure-mode inspection, and released code plus a live demo. The weather-aware net-cost formulation and the safety-oriented constraint evaluation are particularly valuable for retail dynamic-tariff settings with PV. The annualized savings headline is weaker than the controlled scheduling results and should not be the primary takeaway.

major comments (3)
  1. Abstract, §V-E1, Table VII, and §VI-D: the headline claim of approximately £1,270 annual savings is obtained by scaling one archived week (3 passes × 7 days = 21 day-level observations) via 28× mean daily cost and then annualizing. The reported 95% bootstrap CIs for four-week agent cost are extremely wide (e.g., GPT-4o-mini £98.5 with CI £53.8–138.2), and §VII already flags seasonal and price-regime non-representativeness. This projection is not load-bearing for the function-calling or constraint-conflict claims, but it is presented as a primary abstract result. Please demote or strongly qualify the annual figure (report week-level savings share as primary; state that annualization is illustrative only), or extend the archived horizon so the projection is statistically defensible.
  2. §VI-B / Table V / Fig. 3 (S2 and S5): GPT-4o-mini’s 0% success under power-cap and instruction–calendar conflicts persists under both baseline and guided prompts, while Claude is perfect on these families. The paper correctly treats this as a safety property and proposes a deterministic feasibility validator (§VII-B). For the central deployment claim—that commercial agents are practical HEMS interfaces—please make explicit whether the recommended production architecture is (i) Claude-class backends only, (ii) any backend plus mandatory post-hoc validation that can reject/repair schedules, or (iii) a hybrid. Without that, the 100% success / near-MILP optimality narrative from Experiment 1 overstates deployable reliability under realistic household constraints.
  3. §V-D / Table VI and §VI-C: weather-aware co-optimization is reported as regime-dependent and, on sunny days, slightly worse than price-only (+£0.146/day) because negative Agile prices reverse the self-consumption incentive. This is an important and well-observed result, but the abstract still frames weather integration as a core positive contribution without the sunny-day caveat. Please align the abstract and conclusion with the regime-dependent finding (enable weather tools selectively; do not claim unconditional net-cost improvement), so the contribution is not oversold relative to the data.

Circularity Check

0 steps flagged

No circularity: agent schedules are scored against an independently formulated MILP oracle and external rule-based baselines, not against fitted or self-defined quantities.

full rationale

This is an empirical benchmarking paper, not a first-principles derivation. The load-bearing claims (100% multi-appliance success under native function calling; near-MILP optimality; 96.7–98.0% of oracle savings over a week) are obtained by running commercial LLM agents with fixed tools and prompts, then scoring committed schedules against a mixed-integer linear program (Eq. 1–5, Appendix A) that is formulated independently of the agents and solved with PuLP/CBC. Savings shares are defined relative to external zero-inference baselines (off-peak timer, immediate start, greedy cheapest-slot), not relative to quantities fitted from the agent runs. Appliance parameters (Table III), PV model coefficients (Eq. 6: P_stc, α, PR, NOCT), export rate F, and base load B_t are fixed inputs, not estimated to make agents look good. There is no self-citation chain: the single author does not invoke a prior uniqueness theorem or ansatz of their own as external fact; the main literature anchor [17] is by different authors and is used for positioning, not as a load-bearing uniqueness result. The annual ~£1,270 projection is a 28× scaling of one archived week with wide bootstrap CIs—an external-validity limitation the paper itself flags in Sec. VII—not a circular reduction of a prediction to its inputs. Treating net-cost J as the evaluation objective is a construct-validity choice, not circularity of the derivation chain. Score 0 is therefore the honest finding.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 2 invented entities

The central empirical claims rest on standard optimization and agent scaffolding plus a fixed household/tariff/PV configuration. No new physical entities are postulated. Load-bearing free choices are appliance ratings, PV conversion constants, export rate, power cap, and the one-week sample used for annualization. Domain assumptions include treating the linearized net-cost MILP as ground truth and treating tool-committed schedules (not prose) as the only scored actions.

free parameters (7)
  • PV nameplate capacity P_stc = 4.0 kWp
    Fixed at 4.0 kWp in the irradiance-to-generation model (Eq. 6); sets absolute solar surplus available to the net-cost objective and weather-aware arm.
  • PV performance ratio PR and temperature coefficient α = PR=0.80; α=0.004 /K
    PR=0.80 and α=0.004 K^-1 chosen as typical engineering values; they scale forecast generation and thus weather-aware schedules and SCR.
  • NOCT cell-temperature model constant = 45 °C
    NOCT=45°C in the cell-temperature approximation feeding Eq. 6; hand-chosen standard module parameter.
  • Flat export price F = 5 p/kWh
    F≡5 p/kWh as a typical Smart Export Guarantee rate; enters objective (1)/(9) and complementarity cases when C_t < F.
  • Household power cap P_cap = 9 kW
    Set to 9 kW in S2 so EV (7.4 kW) cannot run with another appliance; defines the binding safety conflict used to rank models.
  • Appliance power and duration specs = Table III values
    WM 2.0 kW/4 slots, DW 1.8 kW/3 slots, EV 7.4 kW/12 slots (Table III), following prior work; determine feasible windows and coupling under the cap.
  • Evaluation-week sample for annual projection = 1 week → ×28 / ×52-style annualization
    Seven archived days × 3 passes projected to 28-day and ~£1,270/year savings; the sample choice is free relative to a full-year distribution.
axioms (5)
  • domain assumption Net electricity cost after PV self-consumption (Eq. 1), minimized by the extended MILP, is the correct optimality criterion for scoring agent schedules.
    Section IV defines J and all optimality/gap metrics relative to this objective; comfort, noise, and habit are acknowledged only as construct-validity limits.
  • ad hoc to paper Only the appliance-scheduling tool commit counts as a scheduling decision; fluent prose recommendations are non-executing.
    Section III-B scorer rule; enables the conversational short-circuiting failure mode diagnosis and success-rate definition.
  • domain assumption Native commercial function-calling APIs return structured tool invocations that can be trusted as the agent action interface.
    Experiment 1 contrasts FC vs text-parsed ReAct; H1 and headline 100% success rest on this interface working as specified.
  • domain assumption Pinned evaluation date/time fully grounds relative language (‘tomorrow’) so all models face identical temporal frames.
    Section III-D temporal grounding protocol; without it cross-model comparisons would be confounded.
  • standard math Standard ReAct/tool-calling agent loop and MILP solvability of the daily instance (CBC) are valid computational background.
    Architecture (LangGraph ReAct) and Appendix A formulation; instances solve in <1 s as stated.
invented entities (2)
  • Constraint-conflict scenario suite S1–S6 no independent evidence
    purpose: Reusable adversarial tests where cost optima violate deadlines, power caps, calendars, or feasibility, plus tool-failure recovery.
    Introduced as contribution C2; not a physical entity but a methodological construct. Independent evidence is limited to this paper’s runs unless reused by others.
  • Conversational short-circuiting failure mode independent evidence
    purpose: Name the pattern where a model returns a fluent schedule confirmation with zero tool commits.
    Descriptive label for observed Claude text-interface failures (Sec. VI-A); useful taxonomy, not a new mechanism postulated to explain data.

pith-pipeline@v1.1.0-grok45 · 25638 in / 4138 out tokens · 39438 ms · 2026-07-11T17:06:55.036672+00:00 · methodology

0 comments
read the original abstract

Home Energy Management Systems (HEMS) can reduce residential electricity costs and support demand response, but adoption is limited by the difficulty of translating household preferences into technical scheduling constraints. This paper evaluates whether large language model (LLM) agents can provide a practical natural-language interface for multi-appliance home energy scheduling. We present a tool-calling ReAct agent that uses live half-hourly Octopus Agile prices, weather forecasts, photovoltaic generation estimates, household usage data, and a retrieval-augmented knowledge base to schedule flexible loads against a mixed-integer linear programming (MILP) ground truth. Three commercial models, GPT-4o-mini, Gemini 2.5 Flash, and Claude Sonnet 4.6, are benchmarked across tariff days, constraint-conflict scenarios, weather-aware solar co-optimization, and week-long deployment. With native function calling, all models achieve 100% scheduling success and near-MILP optimality, while text-parsed action interfaces sharply reduce reliability. Constraint testing shows that cost-optimal and safety-optimal models differ: Claude is strongest under infeasibility and power-cap conflicts, while GPT-4o-mini is most efficient. Over a simulated week, agents capture 96.7-98.0% of oracle savings, projecting approximately GBP 1,270 annual savings over an off-peak timer baseline. Code and a live demonstration are available at https://github.com/sokistar24/ecohome-energy-agent and https://www.ecohomeagent.com/.

Figures

Figures reproduced from arXiv: 2607.04569 by Sokipriala Jonah.

Figure 1
Figure 1. Figure 1: System architecture. A LangGraph ReAct agent reasons over a tool layer spanning live external APIs, including the Octopus Energy Agile tariff and [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Cost, latency, and optimality trade-off across the three function-calling [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: It does so by shifting daytime-flexible loads into the [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Constraint-conflict failure rates by scenario family and model under the baseline prompt and the guided prompt. Cell values show failure rate. For [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustrative overcast day (2026-04-26) showing how weather-aware and price-only scheduling place the same three appliances differently. Top: half [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity of weather-aware scheduling to forecast error. The accurate [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Cumulative realized net cost over the seven archived deployment [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Web interface of the EcoHome Energy Agent. The side panel shows the live Agile price profile and household usage and solar patterns. Example [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Thermostat scheduling interaction. The agent retrieves Agile prices, identifies the cheapest and most expensive hours, quantifies the per-kWh saving, [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 1 linked inside Pith

  1. [1]

    Electricity grids and secure energy transitions: Enhancing the foundations of resilient, sustainable and affordable power systems,

    International Energy Agency, “Electricity grids and secure energy transitions: Enhancing the foundations of resilient, sustainable and affordable power systems,” International Energy Agency, Paris, France, Tech. Rep., 2023. [Online]. Available: https://www.iea.org/reports/ electricity-grids-and-secure-energy-transitions

    2023

  2. [2]

    The demand-side flexibility in liber- alised power market: A review of current market design and objectives,

    C. Cabot and M. Villavicencio, “The demand-side flexibility in liber- alised power market: A review of current market design and objectives,” Renewable and Sustainable Energy Reviews, vol. 201, p. 114643, 2024

    2024

  3. [3]

    Demand response,

    International Energy Agency, “Demand response,” 2023, accessed: 2026-07-05. [Online]. Available: https://www.iea.org/energy-system/ energy-efficiency-and-demand/demand-response

    2023

  4. [4]

    Characterizing regional building energy consumption under joint climatic and socioeconomic impacts,

    H. Duan, S. Chen, and J. Song, “Characterizing regional building energy consumption under joint climatic and socioeconomic impacts,”Energy, vol. 245, p. 123290, 2022

    2022

  5. [5]

    Advanced energy management for residential buildings optimizing costs and efficiency through thermal energy storage and predictive control,

    R. Felez and J. Felez, “Advanced energy management for residential buildings optimizing costs and efficiency through thermal energy storage and predictive control,”Applied Sciences, vol. 15, no. 2, p. 880, 2025

    2025

  6. [6]

    Demand- side flexibility in power systems, structure, opportunities, and objectives: A review for residential sector,

    H. Golmohamadi, S. Golestan, R. Sinha, and B. Bak-Jensen, “Demand- side flexibility in power systems, structure, opportunities, and objectives: A review for residential sector,”Energies, vol. 17, no. 18, p. 4670, 2024

    2024

  7. [7]

    Exploiting demand-side flexibility: State-of-the-art, open issues and social perspective,

    F. D’Ettorre, M. Banaei, R. Ebrahimy, S. A. Pourmousavi, E. Blomgren, J. Kowalski, Z. Bohdanowicz, B. Łopaciuk-Gonczaryk, C. Biele, and H. Madsen, “Exploiting demand-side flexibility: State-of-the-art, open issues and social perspective,”Renewable and Sustainable Energy Re- views, vol. 165, p. 112605, 2022

    2022

  8. [8]

    Evaluating decarbonization scenarios and energy management requirement for the residential sector in japan through bottom-up simulations of energy end- use demand in 2050,

    Y . Shimoda, M. Sugiyama, R. Nishimoto, and T. Momonoki, “Evaluating decarbonization scenarios and energy management requirement for the residential sector in japan through bottom-up simulations of energy end- use demand in 2050,”Applied Energy, vol. 303, p. 117510, 2021. 15 TABLE VIII CONSTRAINT-CONFLICTSCENARIOMATRIX ID Family Construction Correct behav...

    2050

  9. [9]

    Barriers to the adoption of energy management systems in residential buildings,

    T. Khafiso, C. Aigbavboa, and S. A. Adekunle, “Barriers to the adoption of energy management systems in residential buildings,”Facilities, vol. 42, no. 15/16, pp. 107–125, 2024

    2024

  10. [10]

    Large language model interface for home energy management systems,

    F. Michelon, Y . Zhou, and T. Morstyn, “Large language model interface for home energy management systems,” inProceedings of the 16th ACM International Conference on Future and Sustainable Energy Systems (e- Energy), 2025, pp. 590–602

    2025

  11. [11]

    Language models are few-shot learners,

    T. Brownet al., “Language models are few-shot learners,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 1877–1901

    2020

  12. [12]

    Improving knowledge extraction from llms for task learning through agent analysis,

    J. R. Kirk, R. E. Wray, P. Lindes, and J. E. Laird, “Improving knowledge extraction from llms for task learning through agent analysis,” in Proceedings of the AAAI conference on artificial intelligence, vol. 38, no. 16, 2024, pp. 18 390–18 398

    2024

  13. [13]

    Theagentcompany: benchmarking llm agents on consequential real world tasks,

    F. F. Xu, Y . Song, B. Li, Y . Tang, K. Jain, M. Bao, Z. Wang, X. Zhou, Z. Guo, M. Caoet al., “Theagentcompany: benchmarking llm agents on consequential real world tasks,”Advances in Neural Information Processing Systems, vol. 38, 2026

    2026

  14. [14]

    ReAct: Synergizing reasoning and acting in language models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao, “ReAct: Synergizing reasoning and acting in language models,” inPro- ceedings of the International Conference on Learning Representations (ICLR), 2023

    2023

  15. [15]

    Rest meets react: Self-improvement for multi-step reasoning llm agent,

    R. Aksitov, S. Miryoosefi, Z. Li, D. Li, S. Babayan, K. Kopparapu, Z. Fisher, R. Guo, S. Prakash, P. Srinivasanet al., “Rest meets react: Self-improvement for multi-step reasoning llm agent,”arXiv preprint arXiv:2312.10003, 2023

    Pith/arXiv arXiv 2023

  16. [16]

    Multi-agent home energy management assistant (hema),

    W. Jung, “Multi-agent home energy management assistant (hema),” SoftwareX, vol. 34, p. 102633, 2026

    2026

  17. [17]

    Agentic AI home energy management system: A large language model framework for residential load scheduling,

    R. El Makroum, S. Zwickl-Bernhard, and L. Kranzl, “Agentic AI home energy management system: A large language model framework for residential load scheduling,”Results in Engineering, vol. 29, p. 109857, 2026

    2026

  18. [18]

    Cost-optimal residential energy scheduling via a zero-shot llm agent and model predictive control,

    V . S. Raghavan and A. Giridhar, “Cost-optimal residential energy scheduling via a zero-shot llm agent and model predictive control,” Journal of Building Performance Simulation, vol. 19, no. 3, pp. 492–510, 2026

    2026

  19. [19]

    Large language model interface for home energy management systems,

    F. Michelon, Y . Zhou, and T. Morstyn, “Large language model interface for home energy management systems,” inProceedings of the 16th ACM International Conference on Future and Sustainable Energy Systems, 2025, pp. 590–602

    2025

  20. [20]

    Llm-mac: Llm-driven multi-agent coordination for generalizable home energy management,

    Y . Yang, Q. Yang, D. Li, L. Sun, L. Sen, and R. Liu, “Llm-mac: Llm-driven multi-agent coordination for generalizable home energy management,” in2025 44th Chinese Control Conference (CCC). IEEE, 2025, pp. 7318–7323

    2025

  21. [21]

    AI agents vs. agentic AI: A conceptual taxonomy, applications and challenges,

    R. Sapkota, K. I. Roumeliotis, and M. Karkee, “AI agents vs. agentic AI: A conceptual taxonomy, applications and challenges,”Information Fusion, vol. 126, p. 103599, 2026

    2026

  22. [22]

    Mcp-enabled agentic ai workflow for building energy modelling: framework and use cases,

    H. Li, L. Zhang, H. Zhou, and T. Hong, “Mcp-enabled agentic ai workflow for building energy modelling: framework and use cases,” Journal of Building Performance Simulation, pp. 1–27, 2026

    2026

  23. [23]

    Exploring llm-based agents for root cause analysis,

    D. Roy, X. Zhang, R. Bhave, C. Bansal, P. Las-Casas, R. Fonseca, and S. Rajmohan, “Exploring llm-based agents for root cause analysis,” in Companion proceedings of the 32nd ACM international conference on the foundations of software engineering, 2024, pp. 208–219

    2024

  24. [24]

    The role of agentic AI in shaping a smart future: A systematic review,

    S. Hosseini and H. Seilani, “The role of agentic AI in shaping a smart future: A systematic review,”Array, vol. 26, p. 100399, 2025

    2025

  25. [25]

    Dyflow: Dynamic workflow framework for agentic reasoning,

    Y . Wang, Z. Xu, Y . Huang, X. Wang, Z. Song, L. Gao, C. Wang, R. Tang, Y . Zhao, A. Cohanet al., “Dyflow: Dynamic workflow framework for agentic reasoning,”Advances in Neural Information Processing Systems, vol. 38, pp. 174 148–174 181, 2026

    2026

  26. [26]

    A survey on agent workflow–status and future,

    C. Yu, Z. Cheng, H. Cui, Y . Gao, Z. Luo, Y . Wang, H. Zheng, and Y . Zhao, “A survey on agent workflow–status and future,” in2025 8th International Conference on Artificial Intelligence and Big Data (ICAIBD). IEEE, 2025, pp. 770–781

    2025

  27. [27]

    A comprehensive review on the applica- tion of large language models (llms) in power systems,

    M. Ghafari, A. Sami, H. Rezapour, S. S. Ghayour, F. Lin, K. Basaran, G. C. Lazaroiu, and P. Siano, “A comprehensive review on the applica- tion of large language models (llms) in power systems,”IEEE Access, vol. 13, pp. 209 450–209 486, 2025

    2025

  28. [28]

    Prompting large language model for multi-location multi-step zero-shot wind power forecasting,

    Z. Duan, C. Bian, S. Yang, and C. Li, “Prompting large language model for multi-location multi-step zero-shot wind power forecasting,”Expert Systems with Applications, vol. 280, p. 127436, 2025

    2025

  29. [29]

    Exploring the capabilities and limitations of large language models in the electric energy sector,

    S. Majumderet al., “Exploring the capabilities and limitations of large language models in the electric energy sector,”Joule, vol. 8, no. 6, pp. 1544–1549, 2024

    2024

  30. [30]

    Llm-empowered decision- focused learning for the operation of local energy communities,

    Y . Zhou, Y . Zuo, D. Kirschen, and Y . Wang, “Llm-empowered decision- focused learning for the operation of local energy communities,”IEEE Transactions on Smart Grid, 2026

    2026

  31. [31]

    Advancing generative artificial intelligence and large language models for demand side management with internet of electric vehicles,

    H. Zhang, R. Zhang, W. Zhang, D. Niyato, Y . Wen, and C. Miao, “Advancing generative artificial intelligence and large language models for demand side management with internet of electric vehicles,” 2025

    2025

  32. [32]

    Ev- stllm: Electric vehicle charging forecasting based on spatio-temporal large language models with multi-frequency and multi-scale information fusion,

    H. Fan, Y . Chai, C. Liu, W. Liu, Z. Zhang, W. Run, and D. Liu, “Ev- stllm: Electric vehicle charging forecasting based on spatio-temporal large language models with multi-frequency and multi-scale information fusion,”Expert Systems with Applications, vol. 313, p. 131620, 2026

    2026

  33. [33]

    Ev charging system considering power dispatching based on multi-agent llms and cgan,

    Z. Niu, J. Li, Q. Ai, J. Jiang, Q. Yang, and H. Zhou, “Ev charging system considering power dispatching based on multi-agent llms and cgan,”IEEE Transactions on Intelligent Transportation Systems, 2025

    2025

  34. [34]

    Dynamic incentive strategies for smart ev charging stations: A llm-driven user digital twin approach,

    Y . Sun, C. Cui, C. Zhang, and C. Gong, “Dynamic incentive strategies for smart ev charging stations: A llm-driven user digital twin approach,” in2025 IEEE International Symposium on the Application of Artificial Intelligence in Electrical Engineering (AAIEE). IEEE, 2025, pp. 17–21

    2025

  35. [35]

    Customized large-scale model for human-AI collaborative operation and maintenance management of building energy systems,

    S. Chen, X. Liang, Y . Liu, X. Li, X. Jin, and Z. Du, “Customized large-scale model for human-AI collaborative operation and maintenance management of building energy systems,”Applied Energy, vol. 393, p. 126169, 2025

    2025

  36. [36]

    Office-in-the-loop: An investigation into agentic AI for advanced building HV AC control systems,

    T. Sawada, M. Mizuno, T. Hasegawa, K. Yokoyama, and M. Kono, “Office-in-the-loop: An investigation into agentic AI for advanced building HV AC control systems,”Data-Centric Engineering, vol. 6, p. e31, 2025. 16 Fig. 7. Web interface of the EcoHome Energy Agent. The side panel shows the live Agile price profile and household usage and solar patterns. Examp...

    2025

  37. [37]

    Can AI make energy retrofit decisions? an evaluation of large language models,

    L. Shu and D. Zhao, “Can AI make energy retrofit decisions? an evaluation of large language models,” 2025

    2025

  38. [38]

    Generating HomeAssistant automations using an LLM-based chatbot,

    M. Giudici, A. Sironi, I. Villa, S. Scherini, and F. Garzotto, “Generating HomeAssistant automations using an LLM-based chatbot,” 2025

    2025

  39. [39]

    Aiot smart home via autonomous llm agents,

    D. Rivkin, F. Hogan, A. Feriani, A. Konar, A. Sigal, X. Liu, and G. Dudek, “Aiot smart home via autonomous llm agents,”IEEE Internet of Things Journal, vol. 12, no. 3, pp. 2458–2472, 2024

    2024

  40. [40]

    AI-driven smart home energy optimization: Integrating AI agents with IoT for adaptive decision- making,

    T.-C. Li, Y .-K. Liu, and Y .-C. Tsai, “AI-driven smart home energy optimization: Integrating AI agents with IoT for adaptive decision- making,” inIET Conference Proceedings, 2025, pp. 225–230

    2025

  41. [41]

    RHEA: Residential home energy advisor,

    N. V . Gkalinikis, C. Nalmpantis, D. Vrakas, S. Chatzigeorgiou, C. Athanasiadis, and D. Doukas, “RHEA: Residential home energy advisor,” inProceedings of the 10th International Conference on Smart and Sustainable Technologies (SpliTech), 2025

    2025

  42. [42]

    Cost-optimal residential energy scheduling via a zero-shot LLM agent and model predictive control,

    V . S. Raghavan and A. V . Giridhar, “Cost-optimal residential energy scheduling via a zero-shot LLM agent and model predictive control,” Journal of Building Performance Simulation, pp. 1–19, 2025

    2025

  43. [43]

    Agile octopus and the octopus energy API,

    Octopus Energy, “Agile octopus and the octopus energy API,” 2026, product AGILE-24-10-01; tariff E-1R-AGILE-24-10-01-C; region C. Accessed: 2026-07-05. [Online]. Available: https://developer.octopus. energy/

    2026

  44. [44]

    Open-meteo.com weather API,

    P. Zippenfenig, “Open-meteo.com weather API,” 2023. [Online]. Available: https://doi.org/10.5281/zenodo.7970649

    doi:10.5281/zenodo.7970649 2023

  45. [45]

    Prompt injection attack against LLM-integrated applica- tions,

    Y . Liuet al., “Prompt injection attack against LLM-integrated applica- tions,” 2023. 17 Fig. 8. Thermostat scheduling interaction. The agent retrieves Agile prices, identifies the cheapest and most expensive hours, quantifies the per-kWh saving, and discloses the electricity price retrieval tool used beneath the response

    2023