Global Optimization and Inference-Time Region Grafting for Agentic Workflows

2026-08-03

AI Summary

Existing agent workflows typically employ a fixed structure determined prior to execution, limiting their ability to dynamically adjust failing components using unlabeled quality signals observed at inference time. This work proposes GRAFT, the first method enabling instance-level adaptation without retraining or global re-optimization. GRAFT leverages unlabeled execution quality signals to evaluate alternative local subroutines and dynamically grafts improved replacements into specific regions while preserving global consistency. The approach is applicable across diverse tasks—including mathematical reasoning, code generation, and multi-hop question answering—and, under identical optimizer and executor configurations, achieves an average performance gain of 3.85 points over the current state-of-the-art method MaAS. Furthermore, GRAFT seamlessly integrates with stronger executors, yielding additional performance improvements.

Abstract

Recent advances in agentic workflow optimization automate workflow design through task-specific workflow search or input-conditioned architecture selection. However, they determine the workflow before execution and cannot adapt failed workflow regions using execution-time label-free quality signals. Naively enabling such inference-time adaptation through whole-workflow re-optimization would be computationally prohibitive. To tackle this challenge, we introduce GRAFT, which preserves a globally optimized workflow while locally replacing only selected regions for each input. Without parameter training, GRAFT evaluates region-level alternatives using label-free execution-quality signals and accepts only replacements that improve local quality while preserving workflow-level consistency, thereby enabling instance-wise adaptation without whole-workflow re-optimization. GRAFT applies without modification across a range of tasks spanning mathematical reasoning, code generation, and multi-hop and knowledge-intensive question answering. Under matched optimizer and executor settings, it improves over the strongest prior workflow-optimization method, MaAS, by 3.85 points on average. Replacing only the executor with a stronger model yields further gains without re-optimizing the global workflow. This suggests that an optimized workflow is not merely a static optimization artifact, but an adaptable execution policy that can evolve with inference-time feedback and stronger executors.

Problem

• agentic workflows
• inference-time adaptation
• workflow optimization
• region grafting
• execution-quality signals

Innovation

• inference-time adaptation
• region grafting
• agentic workflows
• label-free quality signals
• global optimization