Satisficing Agents in Peer-to-Peer ElectricityMarkets: A Compute–Welfare Frontier for Resource-Rational AI

Peer-to-peer (P2P) electricity markets clear every five minutes, leaving little time for complex optimization at the grid edge. We ask a focused question: can lightweight, satisficing agents deliver near-optimizer welfare in continuous double auctions (CDAs) with a fraction of the compute? We build a reproducible agent-based simulator of a residential P2P CDA, instrument per-agent compute, and benchmark an optimizer against two satisficers: an aspiration band (±τ%, whereτ is a price band) and a limited-search rule that inspects at most K offers (agreedy variant accumulates over the first K feasible resting orders; “K-greedy”). On thick markets (N ∈ {200,500}), K-greedy with K ∈ {3,5} attains 100–103% of optimizer normalized welfare while using 40–55×less per-agent compute; results are consistent under a periodic call auction, with a feeder-capacity constraint, and with ticker-only information. Compute scales with offers inspected, and satisficer parameters trace a clear compute–welfare frontier. We measure normalized welfare against a per-interval planner bound and profile compute via per-agent wall-clock time, offers inspected, and peak memory, with instrumentation overhead below 3%. To our knowledge, this is the first quantification of the compute–welfare trade-off for P2P CDAs with explicit per-agent instrumentation and a planner bound for welfare.