Before You Build a Multi-Agent System: An Escalation Framework for LLM Adaptation

Abstract

Multi-agent systems (MAS) have become a popular framework for deploying large language models (LLMs), yet their operational complexity—increased latency, compounding errors, and difficult-to-optimize orchestration—is often unnecessary. Beyond MAS, a rich ecosystem of LLM adaptation strategies exists—spanning input-level methods, parameter updates, and harness based orchestration—yet few principled frameworks guide practitioners who deploy LLMs in real-world systems in choosing among them. To address this gap, we view LLMs as parametric mappings, which makes explicit three adaptation handles ordered by complexity and cost: input level adaptation (X), parameter-level adaptation (θ), and harness-based orchestration (H). Building on this view, we propose an escalation framework that guides practitioners in selecting the most appropriate adaptation strategy for a given task—and, crucially, in exhausting simpler levels before ascending to more complex ones.