KERKIS: a Modeling Language for Multi-Agent System Interactions in AI-Native Software Engineering

Abstract

Agents are increasingly considered first-class participants in the software development lifecycle. However, architectural choices that govern the collaboration of agents are fixed inside the frameworks that implement multi-agent systems, leading to difficulty of measuring the individual contribution of these architectural choices. In this paper, we present a novel modeling language: Knowledge-Explicit Representation of Kinded Interacting Steps (KERKIS). KERKIS can surface architectural choices into independent, manipulable variables. KERKIS represents Actors as execution engines that consume and produce typed Artifacts that flow between Steps and Processes. Using an implementation of KERKIS, we modeled five multi-agent configurations and a single-agent baseline by varying communication mechanisms and coordination strategies. We ran these configurations through a controlled experiment using 60 Infrastructure-as-Code scenarios across four LLMs. Multi-agent configurations modeled in KERKIS significantly outperform the single-agent baseline on non-trivial scenarios. Within the four configurations modeled with KERKIS, orchestration-based coordination consumes 47-112% more tokens than alternatives based on standard operating procedures for a comparable quality of results.