Parallelizing LLM Agent Execution with Contrastive Task Allocation
Experiments on tool-augmented agent workloads show that CoAct improves per-step execution efficiency and resource utilization while achieving competitive or superior task accuracy, demonstrating that contrastive online dispatch can expose substantial parallelism in LLM-agent workflows without retraining the underlying...