AI agents carry electricity-planning scenarios through design, testing and revision

Choosing new power plants, storage and grid upgrades involves more than finding a cheap combination of equipment. A regional investment plan must also work through hourly operation and local electrical constraints. A new preprint explores whether AI agents can manage that repeated cycle of proposing, testing and revising designs.

Electricity planning connects decisions at different scales. Adding renewable generation changes hourly supply; changing battery operation affects local network conditions; reinforcing a line adds cost but may resolve an electrical violation. The preprint Autonomous agents for electricity-system planning, by Yoon Pyo Lee and colleagues, examines whether language-model agents can carry prepared regional scenarios through this process without human direction inside the planning loop.

The researchers report 40 carbon-constrained searches across ten system–year scenarios covering NYISO, MISO, South Africa, Mexico and Brazil. Every search retained a candidate satisfying the framework’s implemented system, representative-feeder and annual operational carbon-budget checks. That is evidence of successful execution within the specified models—not proof that the resulting portfolios are ready to build.

Agents steer; numerical models calculate

The framework divides responsibilities between agents and established modelling tools. Agents choose technology permissions, investment bounds, battery operating conditions and network interventions. PyPSA then optimizes investment and dispatch across 8,760 hourly snapshots. OpenDSS checks the resulting operation on simplified representative distribution circuits, or feeders.

Numerical feedback returns to the agents, which can revise their choices and decide when to stop. A separate reporting record preserves the lowest-cost accepted candidate encountered during the search. The final result therefore need not be the last proposal evaluated.

The annual carbon ceiling is not a promise made by the agent. It is a fixed constraint enforced within the optimization, with a separate acceptance audit. Agents cannot relax it. Similarly, a system-only exploratory calculation cannot establish acceptance: a committed candidate must also pass the applicable feeder checks.

The study tests GPT-6.1 Sol and Claude Opus 5.5, each operating either as a single agent with strategy notes or as a specialist team. Team roles propose experiments, interpret results, criticize decisions and judge convergence. Adaptation happens through conversations, notes and advisory procedures, without task-specific training or changes to model weights.

What the carbon-constrained searches show

The primary evaluation includes selected scenarios for 2023, 2030 and 2035. All four model-and-architecture configurations retained an accepted candidate in every scenario. In 28 of the 40 retained outcomes, emissions reached the scenario ceiling within the reported numerical tolerance.

A separate archive contains 96 searches under a much looser annual ceiling of 1,000 million tonnes of carbon dioxide. Forty match the primary evaluation by system, year, model deployment and architecture. Of those reference designs, 28 exceeded the tighter scenario ceilings.

For NYISO in 2035, the retained carbon-constrained annual cost was about $8.439 billion in all four configurations, 23.56% above the corresponding loose-cap outcomes. The authors treat these as observed trajectory comparisons, not a controlled estimate of the carbon ceiling’s isolated effect. Execution settings and histories differ in some comparisons, and each condition has only one trajectory.

The reference archive also illustrates why local checks matter. In one ERCOT 2035 search, removing a proposed feeder upgrade lowered the annual objective by 0.220% but introduced three hours of undervoltage. The team restored the upgrade, while the reporting rule preserved the earlier accepted design. The apparently cheaper alternative was not an accepted saving.

Passing a budget does not reproduce an official plan

The researchers compare their 2035 portfolios with published Brazilian and South African planning pathways. Brazil’s four carbon-constrained outcomes contain 120.19 gigawatts of total wind capacity, compared with 49.55 gigawatts in PDE 2035. South Africa’s outcomes have a more solar-heavy technology balance than the cited IRP schedule.

These differences are descriptive, not evidence that the agents improved on institutional planning. The comparisons share some inputs, including emissions figures, while retaining different resource, retirement, technology and delivery assumptions. Solar categories also differ because the official plans separately identify distributed generation that the model does not represent as a distinct future rollout.

The boundaries of autonomy

Expert work remains essential before the loop begins: preparing regional data, interpreting sources and defining simulator interfaces. The models use zonal transmission and reduced representative feeders, rather than detailed site-level network studies. Acceptance permits a small amount of unmet annual demand.

Each scenario represents one planning year, with selected additions available without a construction, financing or permitting schedule. Future weather uses one climate model and emissions scenario with reconstructed hourly patterns. The emissions accounting covers direct operational carbon dioxide, not lifecycle emissions or full legal compliance with the source policies.

The authors identify controlled expert comparisons, repeated searches and an adaptive-loop ablation as further tests. The present contribution is narrower: an inspectable autonomous planning process that connects proposals, numerical evidence, revisions and retained accepted candidates under explicit constraints.