AgentCo-op: A Retrieval-Based Synthesis Framework for Interoperable Multi-Agent Workflows

This post introduces AgentCo-op, a framework designed to address the challenges of multi-agent workflow design in open science scenarios. The core idea is to use retrieval-based synthesis to combine existing tools and agents via typed artifact handoffs and bounded self-guided local repair, enabling plug-and-play interoperability without extensive adaptation work.

Multi-agent systems are promising for complex tasks but face key challenges in open science:

1. Lack of curated training data (each research problem is unique, no standard datasets).
2. No reliable scalar evaluation metrics (scientific quality dimensions like novelty are hard to quantify).
3. Missing interface standardization (tools/agents from different teams use incompatible formats, requiring heavy adaptation).

AgentCo-op uses retrieval-based synthesis: instead of designing workflows from scratch, it retrieves and combines existing components (skills, tools, agents). A key innovation is typed artifact handoff: each component defines input/output types, and the framework ensures data transfer meets type constraints (e.g., connecting an agent outputting "gene expression matrix" to one needing it as input), enabling plug-and-play of heterogeneous components.

When execution fails, AgentCo-op uses local repair:

1. Locate problem components via execution logs.
2. Modify only affected components (not the entire workflow).
3. Limit repair search scope to avoid combinatorial explosion.
4. Use execution feedback to guide repairs. This is more efficient than global re-search.

Two case studies validate AgentCo-op:

1. Spatial Transcriptomics Analysis: Combined independent spatial transcriptomics agents, gene set enrichment tools, and biology interpretation agents via typed handoffs to form an auditable workflow without component redesign.
2. Single-Cell Multi-Omics Analysis: Built a parallel workflow for RNA/ATAC modalities, using specialized agents for each and an integration agent to merge results.

AgentCo-op was tested on 6 standard benchmarks (code generation, math reasoning, QA):

• Achieved best results on 4 benchmarks and highest average score.
• Reduced computational cost vs. baselines: a. Avoided global search (local repair is more efficient). b. Reused existing components (no retraining). c. Smart scheduling via type system.

Contributions:

1. Shift from training/search to retrieval-based synthesis (effective for data-scarce open scenarios).
2. Type system as implicit component contract (enables heterogeneous combinations).
3. Local repair principle (better for complex systems).

Limitations:

• Dependent on component library quality/coverage.
• Type system requires domain expert input.
• Local repair may miss global solutions.

Future Work: Auto type inference, component learning from execution history, cross-domain workflow migration.

AgentCo-op demonstrates the potential of retrieval-driven synthesis for open multi-agent workflows. It successfully combines independent tools/agents into executable, auditable workflows using typed handoffs and local repair. A key takeaway: for open complex problems, combining existing resources is more practical than building from scratch, offering a new way to integrate AI systems with existing ecosystems.