Rethinking the Value of Multi-Agent Workflow: A Strong Single Agent Baseline

TL;DR

This paper demonstrates that a single LLM agent can faithfully simulate homogeneous multi-agent workflows via multi-turn conversations and shared KV caches, achieving performance comparable to or better than automated homogeneous MAS while substantially reducing inference cost. It introduces OneFlow, a Monte Carlo Tree Search–driven framework with dual meta-LLMs to automatically design compact workflows optimized for single-agent execution. Experimental results across code, math, QA, and planning benchmarks show strong cost-performance advantages, and a pilot on heterogeneous workflows indicates single-agent approaches can rival some auto-designed heterogeneous configurations while highlighting the remaining potential of true heterogeneity. The work positions single-LLM workflow execution as a strong baseline for MAS research and outlines future opportunities in truly heterogeneous agent systems and end-to-end single-agent training for complex workflows.

Abstract

Recent advances in LLM-based multi-agent systems (MAS) show that workflows composed of multiple LLM agents with distinct roles, tools, and communication patterns can outperform single-LLM baselines on complex tasks. However, most frameworks are homogeneous, where all agents share the same base LLM and differ only in prompts, tools, and positions in the workflow. This raises the question of whether such workflows can be simulated by a single agent through multi-turn conversations. We investigate this across seven benchmarks spanning coding, mathematics, general question answering, domain-specific reasoning, and real-world planning and tool use. Our results show that a single agent can reach the performance of homogeneous workflows with an efficiency advantage from KV cache reuse, and can even match the performance of an automatically optimized heterogeneous workflow. Building on this finding, we propose \textbf{OneFlow}, an algorithm that automatically tailors workflows for single-agent execution, reducing inference costs compared to existing automatic multi-agent design frameworks without trading off accuracy. These results position the single-LLM implementation of multi-agent workflows as a strong baseline for MAS research. We also note that single-LLM methods cannot capture heterogeneous workflows due to the lack of KV cache sharing across different LLMs, highlighting future opportunities in developing \textit{truly} heterogeneous multi-agent systems.

Figures (3)

• Figure 1: Sample question--answer pair from a session-based query recommendation task (left), an exemplar agentic workflow to solve it (middle), and its code representation (right). The workflow demonstrates how multiple LLM agents can collaborate to process complex shopping queries through sequential and conditional execution patterns.
• Figure 2: An example to show how OneFlow framework works. The framework employs dual meta-LLMs (One creative workflow designer and one workflow critic) with Monte Carlo Tree Search to automatically design multi-agent workflows that suitable for single-agent execution for complex tasks. The left panel shows how the first five rounds of workflow design and selection process work. The right panel shows how workflow 6 is generated.
• Figure 3: Pareto frontier: single agent executes multi-agent homogeneous workflow vs. original multi-agent workflows on TravelPlanner xie2024travelplanner benchmark. All the workflows are searched by Claude 4 Sonnet. All the workflows are executed by GPT-4o mini.