Toward Agentic Artificial Intelligence (AI) for Disaster Relief Resource Allocation: An Equity-Aware Multi-Agent Negotiation Framework

Disasters routinely outpace the teams responding to them, and the resource allocation problem at the centre of that response is genuinely hard: scarce medical supplies, food, and shelter must be split across sites that differ in severity, reachable only through a road network that can fail mid-response, and that compete for the same trucks and warehouses.Most multi-agent frameworks in the recent agentic-artificial intelligence literature either compare orchestration tools qualitatively or describe a vision-level architecture without reporting how well the resulting allocation performs.We build a disaster relief resource-allocation environment structured after a published benchmark task for disaster relief under uncertainty and implement a four-agent negotiation framework: a Triage Agent that scores urgency, an Equity Agent that watches for sites falling behind, a Logistics Agent that turns those signals into vehicle dispatches, and a Disruption Agent that reacts to road closures or depot loss; each agent's logic is a simple, auditable rule rather than a large language model (LLM) call, a deliberate design choice we justify and revisit in the discussion.We compare this framework against random, round-robin, and greedy-priority baselines over 90 scenario instances, scored on six standard multi-agent planning metrics.The proposed framework raises goal satisfaction by 44.3% over the strongest baseline (Welch's t = 11.58,p < 0.001) and shows similarly large gains in resource utilisation and disruption recovery, while trading away some of the greedy method's strict severity-first ordering, a trade-off we discuss rather than hide.We close with the limitations of a heuristic-driven simulation and a concrete path toward grounding the agents' decision rules in LLM reasoning for field deployment.