The research
I argue from measurements.
I study whether multi-agent LLM systems can autonomously manage home energy against cost, comfort, and grid constraints. Across 108 simulations the answer so far is yes, and the paper is public with a DOI so anyone can check the method.
- 108simulations behind the published result
- 17-49%household energy cost reduction versus baseline
- 17xcheaper: the model that still matched frontier results
- $0.005per day to run the agent on that model
Published Zenodo preprint March 2026
Benchmarking Multi-Agent LLM Architectures for Home Energy Management
Across 108 simulations, multi-agent LLM architectures cut household energy costs 17 to 49% versus baseline control. Three of four models achieved statistically equivalent mean cost reductions above 20%.
108 runs = 4 models (Llama 4 Maverick, DeepSeek-V3, GPT-4.1, Claude Sonnet 4.6) x 3 real US utility tariffs (ComEd hourly real-time, PG&E E-TOU-C time-of-use, SCE TOU-D-4 tiered with demand charge) x 3 household archetypes x 3 random seeds.
- DeepSeek-V3 delivered savings statistically equivalent to frontier models at $0.005 per day in API cost, 7.5x cheaper than GPT-4.1 and 17x cheaper than Claude Sonnet 4.6.
- Tariff complexity emerged as a stronger model selector than household size.
Fig. 01 / Cross-model cost efficiency
API cost per day, USD
Cost multiples shown relative to DeepSeek-V3.
Open the interactive benchmark
BibTeX
@misc{sulmataj2026benchmarking,
author = {Sulmataj, Besnik},
title = {Benchmarking Multi-Agent LLM Architectures for Home Energy Management: Real-World Tariff Validation and Cross-Model Cost-Efficiency Analysis},
year = {2026},
publisher = {Zenodo},
doi = {10.5281/zenodo.19074522},
url = {https://doi.org/10.5281/zenodo.19074522}
}In progress 2023 - 2026
Optimizing Grid Load Management via Multi-Agent Systems
Doctorate in Business Administration, Management Science. The dissertation studies distributed AI architectures that coordinate autonomously across energy nodes, applying that coordination to real-time load balancing and peak-shaving through cooperative agent decision-making. A parallel thread models smart EV charging networks under NEC Article 625 and NFPA 70:2023, tying the grid-optimization work directly to charging infrastructure design.
Multi-Agent Systems
Distributed AI architectures that coordinate autonomously across energy nodes.
Grid Optimization
Real-time load balancing and peak-shaving via cooperative agent decision-making.
EV Infrastructure
Smart charging networks modeled under NEC Article 625 and NFPA 70:2023.
Also published
The rest of the record.
- Agentic AI for HEMS Benchmark
- Strategic Integration of AI in Energy
- AI Voice Agents in Solar Industry
- Electrons and Algorithms
- AI + Smart Grid for EV Charging
- Featured in Process Reporter
Where it shows up in the build
The research and the work are the same argument.
- Zero-token recall: building memory outside the modelA local second brain that answers recall questions before any model runs, a fleet of scheduled agents, and real tools wired in over MCP.
- Forecasting EV charging demand to optimize smart gridsMachine learning demand models and ChargeGuide, the decision layer that turned a forecast into something an operator could act on.
- All six case studies
- Thoughts on where it goes