Domain-Specific Programming For Agentic Workflows

ACL ARR 2025 May Submission5112 Authors

Abstract: While the use of agentic workflows becomes increasingly contextual, the programmability of it remains general (e.g., prompted steps and tools). In enterprise-related applications, such generalty often leads to undesired downgrade of workflow performance due to the mismatch between a business intention and the specific programming construct a workflow uses to implement it. In this paper, we highlighted the idea of domain-specific programming for agentic workflows, i.e., domain-specific languages (DSL) as a programming interface for both human and AI to create new agentic workflows. This path recovers the nature of agentic workflows as a type of program therefore introduces a systematic way to capture both the structure and semantics of it. Well-defined programming optimizations then become naturally compatible, e.g., check and infer "types". We instantiated the idea by a DSL SOPLang and a multi-agent system which automates the generation of agentic workflows (especially in industrial applications). The preliminary evaluation demonstrated the potential of our proposal to scale in real-world domains.

Paper Type: Short

Research Area: Language Modeling

Research Area Keywords: Generation, Language Modeling

Languages Studied: English

Submission Number: 5112