ReCreate: Reasoning and Creating Domain Agents Driven by Experience
ReCreate employs experience-driven, interaction-based scaffold updates to automate domain agent creation, outperforming black-box methods with 5%+ gains.
Key Findings
Methodology
ReCreate integrates three core modules: experience storage/retrieval, reasoning-creating pipeline, and hierarchical updates. It stores detailed task interaction logs, extracts success/failure signals, and maps these to scaffold edits via a reasoning process. The framework leverages large-scale interaction data as transparent evidence, guiding incremental scaffold refinement. The algorithm involves experience-based updates, pattern abstraction, and multi-round iteration, combining knowledge transfer and causal inference to enhance generalization from minimal seeds. This approach bridges the gap between raw interaction data and scaffold design, enabling efficient domain adaptation.
Key Results
- Across 13 benchmarks spanning software engineering, data science, mathematics, and digital assistance, ReCreate outperformed human-designed and existing automated methods, with an average performance increase of over 5%. In data science and math tasks, improvements reached 8-12 percentage points, demonstrating strong adaptability. Starting from minimal seed scaffolds, ReCreate rapidly generated high-quality agents within few iterations, reducing costs by approximately 30%.
- In complex tasks like software testing and scientific computation, ReCreate achieved 8-12% higher success rates than traditional black-box optimization, confirming robustness. Ablation studies verified the importance of experience storage, reasoning, and hierarchical abstraction modules, showing their combined effect on performance and generalization.
- The framework's ability to leverage interaction logs for precise, evidence-based scaffold updates significantly advances the automation of domain-specific agents, reducing manual effort and enabling scalable deployment.
Significance
This work introduces a paradigm shift by transforming interaction experience into a core driver for agent creation, addressing the high costs and limited scalability of manual design. It enhances interpretability and efficiency, paving the way for autonomous agents capable of continuous self-improvement. The approach aligns with industrial needs for low-cost, adaptable AI systems, promising broad impacts in software automation, intelligent interfaces, and autonomous robotics. Its ability to generalize from minimal seeds accelerates deployment in new domains, reducing reliance on expert knowledge and enabling rapid iteration.
Technical Contribution
The paper proposes a novel white-box optimization framework that explicitly models interaction logs as evidence for scaffold updates. It introduces a reasoning-creating pipeline that converts execution traces into actionable edits, coupled with a hierarchical pattern abstraction mechanism to prevent overfitting. The algorithm integrates bi-level optimization, knowledge transfer, and causal inference, supported by a scalable experience storage system. Empirical results demonstrate superior performance over state-of-the-art methods, validating the theoretical and engineering innovations.
Novelty
This is the first work to systematically utilize detailed interaction experience as the primary driver for automatic domain agent creation, moving beyond performance score-based black-box optimization. The integration of reasoning, evidence extraction, and hierarchical pattern abstraction constitutes a new paradigm in agent engineering, enabling rapid, interpretable, and generalizable agent development from minimal seeds. It bridges the gap between human expert intuition and automated learning, setting a new standard for scalable AI agent synthesis.
Limitations
- The method relies heavily on large-scale interaction data, which can be computationally expensive to process, especially in complex environments. Handling noisy or incomplete logs remains challenging, potentially affecting scaffold quality.
- In highly dynamic or novel tasks, the extracted evidence may be insufficient or misleading, leading to suboptimal updates. The approach may require further adaptation for real-time or online learning scenarios.
- While hierarchical abstraction mitigates overfitting, some domain-specific nuances may still be missed, limiting performance in highly specialized tasks. Future work should focus on improving robustness and data efficiency.
Future Work
Future directions include integrating reinforcement learning to enable agents to autonomously explore and refine scaffolds, incorporating multi-modal interaction data for richer evidence, and extending the framework for online, real-time adaptation. Additionally, exploring meta-learning strategies could further accelerate adaptation to new domains, fostering more autonomous and scalable AI systems. Broader deployment in industrial settings, such as autonomous software engineering and intelligent automation, is also envisioned.
AI Executive Summary
In recent years, large language models (LLMs) like GPT-5 and Claude-4 have revolutionized natural language understanding and generation, leading to the development of intelligent agents capable of complex, multi-step tasks. However, most existing agents depend heavily on manually crafted scaffolds—predefined workflows, rules, and tools—requiring significant human effort and domain expertise. This manual process limits scalability and adaptability, especially in open-world or resource-constrained scenarios.
To address this challenge, the paper introduces ReCreate, a novel framework that leverages rich interaction experience to automatically generate and refine domain-specific agents. Unlike traditional methods that optimize solely based on performance metrics, ReCreate adopts a white-box approach, utilizing detailed execution trajectories, logs, and environmental states as concrete evidence for scaffold updates. The core idea is to treat agent creation as an iterative debugging process, where each interaction provides clues about success or failure, guiding targeted modifications.
ReCreate's architecture comprises three main components: an experience storage and retrieval system, a reasoning-creating pipeline, and a hierarchical update mechanism. The experience storage captures comprehensive interaction data, enabling on-demand inspection. The reasoning pipeline analyzes this data to identify key failure points and generate precise scaffold edits. The hierarchical update module abstracts instance-level insights into reusable domain patterns, ensuring generalization and preventing overfitting. This design allows the framework to start from minimal seeds and evolve powerful, domain-adapted agents efficiently.
Extensive experiments across diverse domains—including software engineering, data science, mathematics, and digital assistance—demonstrate ReCreate's superiority over human-designed and existing automated methods. It achieves over 5% average performance gains, with notable improvements in complex tasks, while significantly reducing costs. These results highlight the potential of experience-driven, interpretable agent creation, promising broad industrial applications.
Looking ahead, integrating multi-modal data, online learning, and meta-learning strategies could further enhance ReCreate's capabilities. Its ability to autonomously adapt and improve from interaction experience paves the way for scalable, intelligent automation systems that require minimal human intervention, fundamentally transforming how AI agents are built and deployed in real-world environments.
Deep Dive
Abstract
Large Language Model agents are reshaping the industrial landscape. However, most practical agents remain human-designed because tasks differ widely, making them labor-intensive to build. This situation poses a central question: can we automatically create and adapt domain agents in the wild? While several recent approaches have sought to automate agent creation, they typically treat agent generation as a black-box procedure and rely solely on final performance metrics to guide the process. Such strategies overlook critical evidence explaining why an agent succeeds or fails, and often require high computational costs. To address these limitations, we propose ReCreate, an experience-driven framework for the automatic creation of domain agents. ReCreate systematically leverages agent interaction histories, which provide rich concrete signals on both the causes of success or failure and the avenues for improvement. Specifically, we introduce an agent-as-optimizer paradigm that effectively learns from experience via three key components: (i) an experience storage and retrieval mechanism for on-demand inspection; (ii) a reasoning-creating synergy pipeline that maps execution experience into scaffold edits; and (iii) hierarchical updates that abstract instance-level details into reusable domain patterns. In experiments across diverse domains, ReCreate consistently outperforms human-designed agents and existing automated agent generation methods, even when starting from minimal seed scaffolds.