$A^2Flow:$ Automating Agentic Workflow Generation via Self-Adaptive Abstraction Operators

TL;DR

This work tackles the automation gap in agentic LLM workflows by introducing A2Flow, a fully automated framework that learns self-adaptive abstraction operators and an operator memory mechanism from expert demonstrations to construct reusable workflow building blocks. It presents a three-stage operator extraction pipeline—Case-based Initial Operator Generation, Operator Clustering and Abstraction, and Deep Extraction for Abstract Execution Operators—to derive compact, general execution operators without manual handcrafting. An Operator Memory Mechanism stores historical operator outputs to enrich context and improve node-level decision-making during workflow search, which is performed via an AFLOW-like MCTS-based optimization that jointly optimizes workflow structure and operators. Empirical results across eight benchmarks in code generation, mathematical reasoning, reading comprehension, embodied tasks, and games show consistent improvements over state-of-the-art baselines, including an average performance gain of $2.4\%$, $19.3\%$ gains on embodied tasks, and $37\%$ reduction in resource usage, highlighting the framework's generalization and scalability for automated agentic workflows.

Abstract

Large language models (LLMs) have shown strong potential in automating the design of agentic workflows. However, existing methods still rely heavily on manually predefined operators, limiting generalization and scalability. To address this issue, we propose $A^2Flow$, a fully automated framework for agentic workflow generation based on self-adaptive abstraction operators. $A^2Flow$ employs a three-stage operator extraction process: 1) Case-based Initial Operator Generation: leveraging expert demonstrations and LLM reasoning to generate case-specific operators; 2) Operator Clustering and Preliminary Abstraction: grouping similar operators across tasks to form preliminary abstractions; and 3) Deep Extraction for Abstract Execution Operators: applying long chain-of-thought prompting and multi-path reasoning to derive compact and generalizable execution operators. These operators serve as reusable building blocks for workflow construction without manual predefinition. Furthermore, we enhance node-level workflow search with an operator memory mechanism, which retains historical outputs to enrich context and improve decision-making. Experiments on general and embodied benchmarks show that $A^2Flow$ achieves a 2.4\% and 19.3\% average performance improvement and reduces resource usage by 37\% over state-of-the-art baselines. Homepage:https://github.com/pandawei-ele/A2FLOW

Figures (4)

• Figure 1: Compared with existing methods, A2Flow improves performance by: (1) Self-Adaptive Abstraction Operators module that optimizes the search process and improves efficiency, and (2) Operators Memory Mechanism that leverages enables more context-aware execution and improving workflow search performance.
• Figure 2: Overview of our framework. Left: Our method synthesizes abstract execution operators through a three-phase generation process, leveraging expert data for iterative refinement. Middle: setting a search space composed of nodes, a abstract operator set, and a code representing edge. Right: Illustration of the MCTS-based evolutionary workflow process. Through integrating an Operators Memory Mechanism, A2Flow enhance the workflow search capability at each node.
• Figure 3: Total execution costs for the partitioned DROP test set are shown. A2Flow and AFLOW-generated workflows (execution model) were evaluated using the same model. Legend colors denote the LLM executing each workflow, with exact values provided in Appendix.
• Figure 4: Self-adaptive process of operator generation and automated workflow optimization. The ALFWorld case study illustrates how A2Flow abstracts embodied tasks into executable operators through three-phase generation, then navigates the search space via MCTS to converge on optimal structured agentic workflows.