CRISP: Compressed Reasoning and Inter-Agent Symbolic Protocol for Efficient Multi-Agent LLM Communication

As large language model (LLM) agents increasingly collaborate in multi-agent systems, inter-agent communication becomes a critical bottleneck, consuming substantial tokens, time, and computational resources. We introduce CRISP (Compressed Reasoning and Inter-agent Symbolic Protocol), a structured communication protocol that replaces verbose natural language exchanges between LLM agents with a compact symbolic vocabulary augmented by compressed logic formula. In multi-agent debate scenarios across four established benchmarks (TruthfulQA, Natural Questions, HaluEval, and FEVER), CRISP achieves comparable or superior accuracy to natural language debate while reducing token consumption up to 79% and communication time up to 76%. Our ablation study shows that the hybrid design, combining symbolic tokens with compressed logic formula, outperforms either component in isolation. These results suggest that LLM agents can effectively communicate through structured symbolic protocols, opening new directions for efficient agentic AI systems.