Optimizing latency in multi-agent systems

Sammanfattning

Large Language Model (LLM)-based multi-agent systems are increasingly used tosolve complex tasks through collaboration between specialized agents. However,the use of multiple agents, tool invocations, and inter-agent communication canintroduce significant latency and cost, limiting practical deployment.This thesis investigates how architectural optimizations affect the performance of anLLM-based multi-agent system. A financial analysis pipeline was implemented usingthe Agent-to-Agent (A2A) protocol for inter-agent communication and the ModelContext Protocol (MCP) for tool use. Four cumulative optimization techniqueswere evaluated: agent parallelization, tool batching, schema pruning, and modelassignment. Performance was assessed using end-to-end latency, inference cost, andoutput quality.The results show that agent parallelization provides negligible latency improvementunder the evaluated deployment conditions due to shared model endpoint contention.In contrast, tool batching reduces median latency up to 27.4% and inference costby 54.6% while improving output quality from 4.18 to 5.00. Schema pruning andmodel assignment techniques further reduces inference cost up to 77.1% comparedto the baseline without degrading quality. Overall, the results suggest that reducingtool invocation overhead and unnecessary context transfer provides greater benefitsthan agent-level parallelization in the evaluated multi-agent architecture.