Truthful and Cost-Minimizing Model Routing in Graph-Based Agentic Workflows

ACL ARR 2026 January Submission6692 Authors

Keywords: Agentic Workflows, Large Language Models, Mechanism Design, Model Routing

Abstract:

As Large Language Models (LLMs) evolve into modular agentic workflows, statically assigning frontier models to every sub-task becomes economically unsustainable. While dynamic routing offers a solution, existing approaches are often myopic: they overlook the substantial context switching costs incurred when transferring state between models, and they fail to account for the strategic incentives of self-interested agents in decentralized marketplaces.
To address these challenges, we propose **STAR** (**S**trategy-proof **T**opology-aware **A**gent **R**outing), a mechanism that simultaneously minimizes execution costs and guarantees truthful bidding. STAR models the allocation problem as a reverse auction on a dependency graph. It employs a virtual cost minimization objective derived from Myerson's optimal auction theory to ensure dominant strategy incentive compatibility (DSIC), while rigorously internalizing the path-dependent externalities of context transfer.
We evaluate STAR across four diverse workflow topologies (e.g., software development, legal analysis) under realistic market volatility. Results demonstrate that STAR significantly reduces total payment compared to baselines.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: multi-agent systems, agent coordination and negotiation, LLM agents, graph-based methods, optimization methods, LLM Efficiency

Contribution Types: Approaches low compute settings-efficiency, Theory

Languages Studied: English

Submission Number: 6692