AgentCo-op auto-assembles multi-agent workflows from existing tools

Designing multi-agent workflows for open-ended scientific tasks is notoriously hard: no curated training sets, no clear evaluation metrics, and tools that don't talk to each other. AgentCo-op, a new framework from researchers, tackles this by retrieving and composing reusable skills, tools, and external agents into executable workflows using typed artifact handoffs. When execution logs reveal a failure, the system performs bounded self-guided local repair on the implicated components—no global topology search or manual redesign needed.

In two real-world genomics case studies, AgentCo-op autonomously orchestrated specialized agents for spatial transcriptomics and gene-set interpretation, enabling collaborative discovery. It also built a parallel workflow for cross-modality marker analysis on single-cell multiome data. The framework even allows importing a searched workflow as a structural prior and improving it via retrieval-based grounding and local repair. On six coding, math, and question-answering benchmarks, AgentCo-op achieved the best result on four and the best average score under a unified backbone, while reducing per-task cost compared to standard multi-agent baselines. The work suggests that retrieval-based synthesis can extend automated workflow design beyond benchmark-optimized graphs to open-world, composed-from-existing-agents scenarios.