Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning to Recommend Multi-Agent Subgraphs from Calling Trees

Xinyuan Song, Liang Zhao

TL;DR

This work tackles the challenge of recommending executable multi-agent configurations from large agent marketplaces by recasting agent selection as a constrained decision problem. It introduces a two-stage framework that first constructs a learned-feasibility set from a historical calling-tree graph and then optimizes a unified scoring function within that set, supporting both single-agent (SARL) and agent-system (ASRL) recommendations. The approach is grounded in a unified calling-tree benchmark built from eight heterogeneous corpora, and it demonstrates that graph-level recommendations yield more coherent and reliable agent teams than agent-level methods, with near-perfect performance on large-scale datasets. The proposed method advances MAS orchestration by incorporating structured interaction effects, historical reliability, and coordination patterns into learning-to-rank, enabling scalable, interpretable, and more dependable multi-agent execution plans.

Abstract

Multi-agent systems (MAS) increasingly solve complex tasks by orchestrating agents and tools selected from rapidly growing marketplaces. As these marketplaces expand, many candidates become functionally overlapping, making selection not just a retrieval problem: beyond filtering relevant agents, an orchestrator must choose options that are reliable, compatible with the current execution context, and able to cooperate with other selected agents. Existing recommender systems -- largely built for item-level ranking from flat user-item logs -- do not directly address the structured, sequential, and interaction-dependent nature of agent orchestration. We address this gap by \textbf{formulating agent recommendation in MAS as a constrained decision problem} and introducing a generic \textbf{constrained recommendation framework} that first uses retrieval to build a compact candidate set conditioned on the current subtask and context, and then performs \textbf{utility optimization} within this feasible set using a learned scorer that accounts for relevance, reliability, and interaction effects. We ground both the formulation and learning signals in \textbf{historical calling trees}, which capture the execution structure of MAS (parent-child calls, branching dependencies, and local cooperation patterns) beyond what flat logs provide. The framework supports two complementary settings: \textbf{agent-level recommendation} (select the next agent/tool) and \textbf{system-level recommendation} (select a small, connected agent team/subgraph for coordinated execution). To enable systematic evaluation, we construct a unified calling-tree benchmark by normalizing invocation logs from eight heterogeneous multi-agent corpora into a shared structured representation.

Learning to Recommend Multi-Agent Subgraphs from Calling Trees

TL;DR

This work tackles the challenge of recommending executable multi-agent configurations from large agent marketplaces by recasting agent selection as a constrained decision problem. It introduces a two-stage framework that first constructs a learned-feasibility set from a historical calling-tree graph and then optimizes a unified scoring function within that set, supporting both single-agent (SARL) and agent-system (ASRL) recommendations. The approach is grounded in a unified calling-tree benchmark built from eight heterogeneous corpora, and it demonstrates that graph-level recommendations yield more coherent and reliable agent teams than agent-level methods, with near-perfect performance on large-scale datasets. The proposed method advances MAS orchestration by incorporating structured interaction effects, historical reliability, and coordination patterns into learning-to-rank, enabling scalable, interpretable, and more dependable multi-agent execution plans.

Abstract

Multi-agent systems (MAS) increasingly solve complex tasks by orchestrating agents and tools selected from rapidly growing marketplaces. As these marketplaces expand, many candidates become functionally overlapping, making selection not just a retrieval problem: beyond filtering relevant agents, an orchestrator must choose options that are reliable, compatible with the current execution context, and able to cooperate with other selected agents. Existing recommender systems -- largely built for item-level ranking from flat user-item logs -- do not directly address the structured, sequential, and interaction-dependent nature of agent orchestration. We address this gap by \textbf{formulating agent recommendation in MAS as a constrained decision problem} and introducing a generic \textbf{constrained recommendation framework} that first uses retrieval to build a compact candidate set conditioned on the current subtask and context, and then performs \textbf{utility optimization} within this feasible set using a learned scorer that accounts for relevance, reliability, and interaction effects. We ground both the formulation and learning signals in \textbf{historical calling trees}, which capture the execution structure of MAS (parent-child calls, branching dependencies, and local cooperation patterns) beyond what flat logs provide. The framework supports two complementary settings: \textbf{agent-level recommendation} (select the next agent/tool) and \textbf{system-level recommendation} (select a small, connected agent team/subgraph for coordinated execution). To enable systematic evaluation, we construct a unified calling-tree benchmark by normalizing invocation logs from eight heterogeneous multi-agent corpora into a shared structured representation.
Paper Structure (53 sections, 24 equations, 4 figures, 7 tables)

This paper contains 53 sections, 24 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Methodology Overview
  3. Problem Formulation
  4. SARL: Single-Agent Recommendation Learning
  5. ASRL: Agent-System Recommendation Learning
  6. Method
  7. Single Agent Recommendation learning (SARL)
  8. SARL Retrieval.
  9. SARL Rerank and Recommendation.
  10. Agent-system Recommendation Learning (ASRL)
  11. ASRL Retrieval.
  12. ASRL Rerank and Recommendation.
  13. Token Complexity Analysis.
  14. Token complexity of Naive ranking.
  15. Token complexity of our SARL.
  16. ...and 38 more sections

Figures (4)

  • Figure 1: Agent marketplace and agent-system growth trends in 2025. (a) Number of marketplace agents per quarter. (b) Number of marketplace categories per quarter. (c) Average number of agents per system per quarter. All panels report quarterly snapshots in 2025 (Q1--Q4).
  • Figure 2: Overview of our two-stage constrained recommendation framework for multi-agent orchestration. (1) Inputs & data foundation: a task query, a unified calling-tree dataset, and a global agent network for assisting calling-tree construction. (2) Core methodology (SARL/ASRL): retrieve feasible agents/subgraphs under learned constraints, then rank with a unified scorer (relevance, history, cooperation, structure). (3) Outputs & targets: recommended agents/agent systems and a calling-graph plan, targeting success, reliability, cooperation accuracy, and structural coherence.
  • Figure 3: Learning-to-rank performance comparison: SARL vs. multi-agent ASRL.
  • Figure 4: Unified multi-agent dataset representation, including calling-tree structure and agent pool metadata. Nodes represent single tool/API invocations, and edges encode execution ordering and dependency relations.