[2E1-GS-5b] Agents

9:00 AM - 9:15 AM JST(12:00 AM - 12:15 AM UTC)

[2E1-GS-5b-01] Agentic Flow: A Framework for Explicit Agent Workflow DescriptionA Framework for Explicit Agent Workflow Description

〇Takuma Shibahara¹, Shogo Suga¹, Shotaro Medera¹, Tomohiro Onoue¹ (1. Daiichi Sankyo Company, Limited)

Keywords:

Agent,Agentic AI,AI Agent

Multi-agent workflows based on large language models suffer from implicit state updates, boilerplate overhead, and hidden state transitions, making code difficult to understand and maintain. This study proposes Agentic Flow (AF), a framework that separates agent specification from execution. In AF, agent calls return specification objects (ExecutionSpec) without triggering execution; execution occurs only at explicit await points, enabling workflows to be written in plain Python with standard control flow constructs. The framework further separates concerns into five orthogonal axes and provides phase-based context management. In comparative experiments with three multi-agent workflow tasks, AF achieved up to 42% code reduction while preserving parallel execution capability during streaming–a property that existing frameworks sacrifice. Exception injection tests confirmed unique identifiability of execution points and consistent pairing of state mutation boundaries. These results demonstrate that the proposed design improves both readability and maintainability of agent workflow implementations.