CEDAR: Carbon Efficient Dynamic Allocation and Routing for Agentic LLM Inference

LLM inference now dominates operational AI compute, yet production serving stacks typically optimise for performance alone, leaving cost and carbon unmanaged. We present CEDAR, a queue level multi objective control framework for agentic LLM inference that jointly optimises tail latency, cloud cost, and marginal carbon emissions. CEDAR observes backlog depth, waiting time percentiles, and service level objective (SLO) slack to route mixed criticality requests across heterogeneous, geo-distributed fleets. In trace-driven evaluation, CEDAR reduces cost by up to 26% and carbon by up to 27% relative to a Performance-Only baseline, while maintaining competitive p95 latency (0.88 s) and low SLO violation (4.3%). These results indicate queue level control as a practical path to sustainable agentic inference without unacceptable QoS degradation.