LLM Microgrid Agents
Can LLM agents, one per household, negotiate in plain English to fairly share limited solar and battery power during a grid outage? Yes: live agents beat a zero-LLM control on every clean seed, and fairness improved alongside.
Overview
Climate disasters cause long grid outages, and households ride them out very unequally: some have rooftop solar and large batteries, others only small batteries. Classical optimization can allocate the neighborhood's energy fairly, but it assumes a central controller with perfect, pre-formalized information, and it cannot explain its decisions to the residents who live with them.
This research asks whether a population of LLM agents, one per household, can close that gap by negotiating peer-to-peer in natural language: reporting needs, updating beliefs about neighbors from nothing but what peers say, making binding commitments to transfer energy, and justifying their actions afterward. The agents work with limited and potentially inaccurate information, just as real neighbors would.
On the three clean seeds of the showcase scenario, live agents raised total served load by 5.8 ± 1.0 percentage points over a zero-LLM control, closing 29.0% ± 2.9 of the gap between that control and a perfect-information oracle, with distributional fairness improving at the same time: no efficiency-versus-equity tradeoff. An LLM judge distinct from the agent model found the agents' explanations actionable and consistent with their actions, though weaker on state accuracy. The stress tests below show where the approach bends and where it breaks.
Natural-language negotiation added +5.8 ± 1.0 points of served load over a zero-LLM control, on all 3 of 3 seeds
Earlier naive versions of the agent layer lost to simple heuristics; the result flipped only after reworking how information flows, so that beliefs come solely from peer reports and negotiated commitments actually bind. The comparison is against a control that runs identical machinery with deterministic policies, so the delta isolates what the LLM itself adds.
Methodology
Deterministic Simulator
A 30-household neighborhood (5×6 grid) simulated in 15-minute ticks across a 24-hour outage, with realistic battery physics (90% round-trip efficiency, depth-of-discharge floor, transit losses, a shared bus cap) and synthetic load and solar profiles. Fully deterministic: every number regenerates at $0 from committed caches.
LLM Agent Layer
One Claude agent per household. Agents communicate through typed natural-language messages, update beliefs about neighbors solely from what peers report, perceive their own state through a noisy interface, and establish binding commitments to transfer energy.
Controls & Baselines
The key comparison: a zero-LLM control that runs the identical machinery with deterministic policies in place of the LLM. Upper bound: a linear-programming oracle with perfect information. A round-robin heuristic is reported as a secondary reference.
Failure-Axis Stress Tests
Beyond clean runs, single-seed cells probe robustness: defecting households that withhold generation, observation noise on battery and load readings, and a strict cap on how many messages agents may send per tick.
Robustness Stress Tests
Defectors
With 33.6% of generation withheld by defecting households, live agents still beat the control (single seed), retaining 89.1% of coordination value where naive proportional scaling predicts 66.4%.
Observation Noise
With ±10% noise on battery state and ±15% on load, the clean-cell edge of +4.6 points compresses to +0.9 (single seed): coordination survives bad information, but barely.
Message Budget
Capped at 2 messages per tick, live agents fall below the control (single seed). Natural-language coordination has real bandwidth overhead; most messages were budget-dropped.