AgentEvo: cost-aware agentic workflow generation via adaptive multi-stage evolution

The automation of agentic workflow generation is increasingly important for deploying LLMs in complex tasks. However, current optimization methods often incur substantial token-based cost due to structural redundancy and token inflation, a phenomenon we term empirical additive bias exhibited by LLM-based optimizers under workflow-improvement optimization prompts. To address this, we propose AgentEvo, a cost-aware framework for agentic workflow generation that balances accuracy and token-based cost without human intervention. Specifically, we introduce (1) an adaptive multi-stage evolution strategy that decouples exploration, refinement, and crossover for redundancy pruning and operator fusion, and (2) a cost-aware parent-selection policy that preserves diverse Pareto-optimal candidates. Experimental evaluations on the MATH, GSM8K, HumanEval, and MBPP benchmarks show that AgentEvo consistently improves the accuracy–cost trade-offs over strong automated workflow optimization baselines. On MATH it attains an accuracy of 52.47% while reducing the average computational cost by 40% relative to the strongest baseline.