Evoflux: Inference-Time Evolution of Executable Tool Workflows for Compact Agents

Researchers introduce Evoflux, an inference-time evolutionary search method that significantly improves how compact language models handle tool use and workflow execution. By treating tool failures as a repair problem rather than a generation problem, Evoflux increases execution feasibility from 3% to 17-24% on complex multi-tool tasks, outperforming traditional fine-tuning approaches while maintaining cost efficiency.

Evoflux addresses a critical gap in deploying smaller language models as autonomous agents. While compact models offer substantial cost and latency advantages over large models, they struggle with the complex orchestration required for real-world tool use—discovering available tools, validating parameters, tracking dependencies, and executing reliable workflows. Traditional approaches rely on fine-tuning with teacher demonstrations, but this method fails to teach agents how to recover from execution failures or adapt to changing tool catalogs, which are inherent aspects of production systems.

The research builds on established trends in AI efficiency and agentic systems. As organizations seek to reduce computational overhead and deployment complexity, compact models become increasingly attractive. However, their limitations in reasoning and planning have confined them to narrow use cases. This work demonstrates that execution-grounded search—allowing models to iteratively refine workflows based on real feedback—overcomes these constraints more effectively than scaling up training data or model size.

The practical implications are substantial. Developers building agent systems can now deploy smaller, faster models without sacrificing reliability on complex multi-tool tasks. The 17-24% feasibility rate represents a meaningful improvement in real-world applicability, particularly for systems with unpredictable tool availability or changing requirements. The method's superiority over supervised fine-tuning and direct preference optimization (DPO) suggests that inference-time computation, not training data alone, determines performance in dynamic environments.

Looking forward, similar repair-and-refine approaches may become standard for autonomous systems. The research hints at broader architectural shifts where runtime search and feedback mechanisms replace static learned patterns for handling edge cases and environmental changes.

• →Evoflux achieves 17-24% execution feasibility on MCP tool tasks, up from 3% for base compact models, through evolutionary search at inference time
• →Evolutionary workflow repair outperforms supervised fine-tuning and DPO approaches despite using identical training data, suggesting inference-time computation matters more than scale
• →The method handles dynamic tool catalogs and execution failures that standard distillation cannot teach through fixed demonstration sets
• →Compact agents with Evoflux can compete with larger models on complex multi-step tool orchestration while maintaining latency and cost advantages
• →Execution-grounded feedback enables adaptive recovery behaviors, meta-guided redesign, and structured workflow evolution for tool use agents