Opening — Why this matters now
The energy transition is quietly becoming a coordination problem.
Solar panels, home batteries, electric vehicles, heat pumps, and smart appliances are turning millions of households into “prosumers” — participants that both consume and produce electricity. In theory, these distributed assets could collectively provide enormous flexibility to stabilize power grids.
In practice, participation remains stubbornly low.
Demand response (DR) programs — where households adjust electricity usage in response to grid needs — rarely achieve more than modest adoption. The reasons are surprisingly mundane: confusing notifications, opaque incentives, and a sense that the system is acting on the household rather than with it.
The paper introduces a deceptively simple idea: what if energy coordination worked more like a conversation?
Instead of silent automation or one‑way price alerts, LLM-powered agents could allow the grid and households to negotiate flexibility in natural language. The authors call this concept Conversational Demand Response (CDR).
If it works, the electricity grid might start to behave less like a command hierarchy — and more like a marketplace with millions of tiny negotiators.
Background — The limits of traditional demand response
Demand response has existed for decades, but most implementations fall into two categories:
| Approach | How it works | Main limitation |
|---|---|---|
| Automated control | Utility directly adjusts devices | Low transparency and trust |
| Price signals | Users respond to dynamic pricing | Requires active attention |
Both approaches scale technically, but neither scales socially.
Fully automated systems reduce cognitive load but remove user agency. Meanwhile, price alerts and dispatch notifications push too much complexity onto households.
Research in behavioral energy economics repeatedly shows three barriers to sustained participation:
| Barrier | Explanation |
|---|---|
| Loss of control | Users fear devices acting unpredictably |
| Lack of transparency | Incentives and grid needs are unclear |
| Cognitive effort | Managing energy manually is tedious |
The result is a paradox: residential assets have enormous theoretical flexibility, yet most of it remains untapped.
The authors propose solving this not with better optimization — but with better interaction.
Architecture — When the grid and the home both run AI agents
The proposed system uses a two‑tier multi‑agent architecture.
At a high level, two autonomous agent systems coordinate with each other:
| Layer | Agent | Role |
|---|---|---|
| Grid level | Aggregator agent | Manages many households and dispatches flexibility requests |
| Household level | HEMS agent | Evaluates requests and negotiates participation |
The household agent operates within a Home Energy Management System (HEMS) that controls devices such as batteries, EV chargers, and appliances.
The interaction model looks surprisingly human:
- The aggregator identifies a grid event (e.g., evening demand spike).
- It asks households for flexibility.
- Each household agent evaluates feasibility.
- The system explains the cost‑benefit to the resident.
- The resident approves or rejects the request.
In other words, the optimization remains technical — but the interface becomes conversational.
Implementation — Optimization inside the conversation
The most interesting engineering element is how mathematical optimization becomes part of the dialogue.
Within the household system, the battery is managed by a specialized sub‑agent that calls a Mixed‑Integer Linear Programming (MILP) optimizer.
The optimization objective balances three competing factors:
$$ \text{Minimize cost} = \sum_t [(p^{imp}_t \cdot \pi_t - p^{exp}t \cdot \pi{fit})\Delta t
- (p^{ch}_t + p^{dis}t)c{deg}\Delta t]
- d_{peak} w_{peak} $$
Where the optimizer considers:
| Variable | Meaning |
|---|---|
| Grid imports/exports | Electricity bought or sold |
| Battery charge/discharge | Energy storage decisions |
| PV generation | Solar production |
| Household demand | Local electricity consumption |
The optimizer performs a dual‑solve procedure:
- Solve the baseline schedule without DR participation
- Solve again with the DR commitment constraint
This comparison allows the system to explain the economics of participation.
For example, a household might receive a message like:
“Providing 3 kW between 17:00–19:00 will earn €1.13. The battery will discharge earlier but recharge later. No comfort impact.”
The key innovation is not the optimization itself — utilities already do that — but the translation of optimization outputs into understandable explanations.
Findings — Does conversational coordination actually work?
The authors implemented a proof‑of‑concept system with LLM‑based agents and measured computational performance across different scenarios.
Performance metrics
| Scenario | Iterations | Tool Calls | Tokens | Avg Time |
|---|---|---|---|---|
| DR acceptance | ~3.6 | ~2.6 | 23k | 8.3 s |
| DR rejection | ~5.0 | ~3.6 | 34k | 9.8 s |
| High target request | ~3.4 | ~2.4 | 21k | 7.8 s |
| Profile updates | 1 | 0 | ~1k | ~1.5 s |
Two observations stand out:
1. Conversations remain fast enough for operational use
Even complex negotiation scenarios complete in under 10 seconds.
2. Optimization remains the computational bottleneck
LLM reasoning cycles are relatively cheap compared to running the energy optimization model.
In other words, AI is not replacing engineering models — it is orchestrating them.
Implications — What conversational infrastructure could unlock
The significance of CDR goes beyond electricity.
It demonstrates a pattern that will likely appear across many infrastructure systems:
AI agents acting as negotiators between humans and complex optimization systems.
Three implications stand out.
1. Energy markets could become conversational
Instead of opaque tariffs and static programs, households could dynamically negotiate flexibility with the grid.
2. Agentic coordination scales distributed infrastructure
Power systems increasingly rely on millions of small assets rather than a few large power plants.
Conversational agents may be the only practical way to coordinate this complexity.
3. Transparency becomes a system feature
Because all decisions must be explainable in natural language, optimization models effectively gain an interpretability layer.
This could improve trust — a critical factor for participation in energy programs.
Conclusion — When infrastructure learns to talk
The most interesting part of this research is philosophical rather than technical.
Power grids historically evolved as centralized command systems. Even “smart grids” still rely heavily on automated dispatch logic and invisible algorithms.
Conversational Demand Response suggests a different future.
In that future, the grid does not simply control devices — it asks.
Your home battery evaluates the request. Your AI energy assistant explains the trade‑off. And you decide whether the grid gets the energy it needs.
The electrons still move through wires.
But the coordination layer becomes something entirely new: a conversation between millions of autonomous agents.
Cognaptus: Automate the Present, Incubate the Future.