Xiang, K., Shi, B., Lyu, Y., Chen, J., Tsang, I., & Yin, H. (2026). MermaidFlow-CF: How Agentic Workflow Representation Governs Constraint-Faithful Control.
Abstract:
Agentic workflows coordinate LLM agents to autonomously generate, refine, and execute multi-step task pipelines, yet maintaining reliable control over instruction-following behavior remains challenging, often resulting in cascading workflow failures. These failures frequently stem from unconstrained workflow synthesis, where structural drift and broken control flow accumulate over time. In this paper, we show that a key driver of this brittleness is workflow representation, which determines whether planning structure and control flow can be preserved during generation, evaluation, and execution. We introduce MermaidFlow-CF, a constraint-faithful workflow optimization framework that represents workflows in a symbolic graph DSL Mermaid, enabling symbolic control-flow syntax that renders planning structure explicit and supports interpretable and checkable workflow generation. Building on this formulation, we formalize the constrained workflow optimization problem and introduce a structured taxonomy of workflow constraints spanning resource feasibility, executability, structural validity, and causal coherence. We further develop an evaluation protocol to measure constraint violation and correction dynamics for the constraints. Across multi-step reasoning benchmarks, MermaidFlow-CF achieves significantly higher constraint fidelity and markedly fewer cascading failures than AFlow, a Python-based workflow optimization baseline. These results show that symbolic workflow representations in Mermaid provide a more reliable foundation for agentic pipelines than Python, and that constraints function not as barriers but as structural priors that shape optimization dynamics and enable more stable, higher-performance optimization in agentic workflow planning.
License type:
Publisher Copyright
Funding Info:
This research / project is supported by the National Research Foundation, Singapore - AI Singapore Programme
Grant Reference no. : AISG-NMLP-2024-003
This research / project is supported by the National Research Foundation, Singapore and Infocomm Media Development Authority - Trust Tech Funding Initiative
Grant Reference no. : DTC-RGC-05
This research / project is supported by the Agency for Science, Technology and Research (A*STAR) - Career Development Fund (CDF)
Grant Reference no. : C233312007
This research / project is supported by the Agency for Science, Technology and Research (A*STAR) - Career Development Fund (CDF)
Grant Reference no. : C243512014