AgentCo-op: Retrieval-Based Synthesis of Interoperable Multi-Agent Workflows

TL;DR

AgentCo-op introduces a retrieval-based framework for synthesizing multi-agent workflows in open-ended scientific settings, enabling composition, repair, and improvement of workflows without extensive redesign.

Contribution

It presents a novel retrieval-based synthesis method that composes and repairs multi-agent workflows from existing components in open-world scenarios.

Findings

Successfully applied to genomics case studies for spatial transcriptomics and gene-set interpretation.

Achieves best results on four out of six benchmarks and improves average scores.

Reduces per-task cost compared to multi-agent baselines.

Abstract

Designing multi-agent workflows is especially difficult in open-ended scientific settings where tasks lack curated training sets, reliable scalar evaluation metrics, and standardized interfaces between existing tools and agents. We propose AgentCo-op, a retrieval-based synthesis framework that composes reusable skills, tools, and external agents into executable workflows through typed artifact handoffs, then applies bounded self-guided local repair to implicated components when execution evidence indicates failure. In two open-world genomics case studies, AgentCo-op composes independently developed scientific agents and external tool repositories into auditable workflows without redesigning them or running global topology search. It coordinates specialized agents for spatial transcriptomics and gene-set interpretation to enable collaborative discovery from spatial transcriptomics data, and builds a parallel workflow for cross-modality marker analysis on single-cell multiome data. AgentCo-op can also import a searched workflow as a structural prior and improve it by grounding nodes with retrieved components and applying local repair, showing that synthesis and search are complementary. On six coding, math, and question-answering benchmarks, AgentCo-op achieves the best result on four benchmarks and the best average score under a unified backbone setting, while consistently reducing per-task cost relative to multi-agent baselines. Together, these results suggest that retrieval-based synthesis can extend automated agentic workflow design beyond benchmark-optimized agent graphs to open-world workflows built from existing agents, tools, and typed artifacts.