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

Case Study 1 · Coordinating Domain Agents (TissueAgent × GeneAgent)

On a developing human heart MERFISH dataset, AgentCo-Op is asked whether aFibro cells in the AVN / AV-ring cellular community exhibit a distinct transcriptional program relative to aFibro cells in the left and right atria. From only a task description and the GitHub URLs of TissueAgent (spatial transcriptomics) and GeneAgent (gene-set interpretation), AgentCo-Op profiles both repositories, builds isolated Docker containers, registers each as an external workflow node, and synthesizes a broker-mediated handoff workflow. The synthesized pipeline identifies 576 target aFibro cells against 5,685 controls, recovers 53 upregulated markers, and the downstream GeneAgent interprets them as an AV-canal- and node-associated fibroblast program — concluding that AVN / AV-ring aFibro cells represent a developmentally specialized, ECM-rich, conduction-niche-associated state.

Case Study 2 · Composing Domain Workflows (Seurat × Signac)

On the 10x PBMC multiome dataset, AgentCo-Op composes Seurat (scRNA-seq) and Signac (scATAC-seq) into a parallel cross-modality marker-discovery workflow with an explicit join step. Markers are evaluated against CellMarker 2.0 and PanglaoDB: the intersection of modalities yields jointly supported markers (evaluated for precision), and the union captures all recovered markers (evaluated for recall). Combining the two modalities improves both macro precision and recall over either modality alone across both reference databases.

Table 1. Macro precision (on the intersection) and recall (on the union) of cross-modality marker integration on the PBMC multiome dataset. The Combined column is the cross-modality result.

Case Study 3 · Reusing Existing Agent Graphs (AFlow → AgentCo-Op)

AgentCo-Op imports the multi-agent graph produced by a trained AFlow search on MBPP and treats it as a structural prior in Ω. AgentCo-Op then resynthesizes the agent graph, grounds its nodes with retrieved skills and tools, and applies bounded local repair during execution. The hybrid AFlow + AgentCo-Op outperforms both AFlow alone and AgentCo-Op built from scratch — showing that synthesis and search are complementary.

Table 2. MBPP performance of different agentic workflow design strategies. Initializing AgentCo-Op from an AFlow-searched graph improves performance compared with initializing it from scratch.

Standard Benchmarks (Matched Backbone: GPT-4o-mini)

Across six benchmarks spanning QA (HotpotQA, DROP), code generation (HumanEval, MBPP) and math reasoning (GSM8K, MATH), AgentCo-Op achieves the best result on 4 / 6 benchmarks and the best average score under matched-backbone conditions — without any workflow search or training stage. AFlow^* denotes results reported in the original AFlow paper (mixed backbones); AFlow (GPT-4o-mini) is our matched-backbone rerun.

Table 3. Performance across six benchmarks using GPT-4o-mini as the backbone. Bold indicates the best score.

Cost Analysis

AgentCo-Op is substantially more efficient than discussion-based multi-agent baselines. It separates one-time workflow synthesis from bounded instance-level repair, avoiding both the training-time search cost of AFlow and the per-instance round-trip cost of LLM-Debate / ReConcile. Test-time cost is lower than ReConcile on all six benchmarks and lower than LLM-Debate on five of six.

Table 4. Per-dataset performance and aggregate cost. A dash in Train Cost indicates that the method requires no workflow search or training stage. Bold = best score per dataset.