AgentFold is presented, a multi-agent framework that formulates folding-model development as a closed-loop search over executable code variants and improves the best lDDT by 7.5% over independent Codex proposals and outperforms a random-search control.
Abstract
Scientific LLM agents have shown promise in literature reasoning, tool use, and experiment planning, but it remains unclear whether they can autonomously improve large, tightly coupled scientific machine-learning systems through executable code changes and computationally expensive validation. We study this question in protein folding, where progress requires coordinated architectural modifications, multi-objective evaluation, and domain-aware interpretation. We present AgentFold, a multi-agent framework that formulates folding-model development as a closed-loop search over executable code variants. Starting from ESMFold, AgentFold proposes hypotheses, implements and debugs code-level modifications, evaluates model variants, analyzes experimental outcomes, and stores both successful and failed interventions in structured memory. An MCTS-style policy allocates computational resources across high-scoring search branches. On an engineering-scale protein-folding codebase comprising more than 2,000 lines of code, AgentFold explores approximately 80 model variants using approximately 5,000 GPU-hours and 170 million LLM tokens. Under a matched computational budget, AgentFold improves the best lDDT by 7.5% over independent Codex proposals and outperforms a random-search control. Beyond model improvement, the resulting intervention traces reveal recurring empirical design patterns: stable gains tend to arise from early, soft, learnable priors and gated refinement, whereas direct geometric perturbations and geometry-conditioned feedback often destabilize training. The code and experimental resources are publicly available at https://github.com/lmqfly/AgentFold.
In the context of cloud computing, risks associated with underlying technologies, risks involving service models and outsourcing, and enterprise readiness have been recognized as potential barriers for the adoption. To accelerate cloud adoption, the concrete barriers negatively influencing the adoption decision need to be identified. Our study aims at understanding the impact of technical and security-related barriers on the organizational decision to adopt the cloud. We analyzed data collected through a web survey of 352 individuals working for enterprises consisting of decision makers as well as employees from other levels within an organization. The comparison of adopter and non-adopter sample reveals three potential adoption inhibitor, security, data privacy, and portability. The result from our logistic regression analysis confirms the criticality of the security concern, which results in an up to 26-fold increase in the non-adoption likelihood. Our study underlines the importance of the technical and security perspectives for research investigating the adoption of technology.
Nattakarn Phaphoom, Xiaofeng Wang, S. Samuel et al.· Journal of Systems and Softw...· 111 citations· ⚡8
To compete in this age of disruption, large companies cannot rely on cost efficiency, lead time reduction and quality improvement. They are now looking for ways to innovate like startups. Meanwhile, the awareness and use of the Lean startup approach have grown rapidly amongst the software startup community in recent years. This study investigates how Lean internal startup facilitates software product innovation in large companies and identifies its enablers and inhibitors. A multiple case study approach is followed in the investigation. Two software product innovation projects from two large companies are examined, using a conceptual framework that is based on the method-in-action framework and extended with the previously developed Lean-Internal Corporate Venture model. Seven face-to-face in-depth interviews of the employees with different roles are conducted. Within-case analysis and cross-case comparison are applied to draw the findings from the cases. A generic process flow summarises the common key processes of Lean internal startups. The findings suggest that an internal startup that is initiated management or employees faces different challenges. A list of enablers of applying Lean startup in large companies are identified, including top management support and cross-functional team. Both cases face different inhibitors due to the different process of inception, objective of the team and type of the product. Our contributions are threefold. First, this study is one of the first attempt to investigate the use of Lean startup approach in large companies empirically. Second, the study shows the potential of the method-in-action framework to investigate the Lean startup approach in non-startup context. The third is a general process of Lean internal startup and the evidence of the enablers and inhibitors of implementing it, which are both theory-informed and empirically grounded.
Henry Edison, Nina M. Smørsgård, Xiaofeng Wang et al.· Journal of Systems and Softw...· 78 citations· ⚡6
Affects--emotions and moods--have an impact on cognitive processing activities and the working performance of individuals. It has been established that software development tasks are undertaken through cognitive processing activities. Therefore, we have proposed to employ psychology theory and measurements in software engineering (SE) research. We have called it "psychoempirical software engineering". However, we found out that existing SE research has often fallen into misconceptions about the affect of developers, lacking in background theory and how to successfully employ psychological measurements in studies. The contribution of this paper is threefold. (1) It highlights the challenges to conduct proper affect-related studies with psychology; (2) it provides a comprehensive literature review in affect theory; and (3) it proposes guidelines for conducting psychoempirical software engineering.
D. Graziotin, Xiaofeng Wang, P. Abrahamsson· SSE@SIGSOFT FSE· 56 citations· ⚡4
It is essential for startups to quickly experiment business ideas by building tangible prototypes and collecting user feedback on them. As prototyping is an inevitable part of learning for early stage software startups, how fast startups can learn depends on how fast they can prototype. Despite of the importance, there is a lack of research about prototyping in software startups. In this study, we aimed at understanding what are factors influencing different types of prototyping activities. We conducted a multiple case study on twenty European software startups. The results are two folds; firstly we propose a prototype-centric learning model in early stage software startups. Secondly, we identify factors occur as barriers but also facilitators for prototyping in early stage software startups. The factors are grouped into (1) artifacts, (2) team competence, (3) collaboration, (4) customer and (5) process dimensions. To speed up a startup’s progress at the early stage, it is important to incorporate the learning objective into a well-defined collaborative approach of prototyping.
Anh Nguyen-Duc, Xiaofeng Wang, P. Abrahamsson· International Conference on...· 44 citations· ⚡5
This study constructed a pH-responsive P-TN/SF@Fe-Cur composite coating that demonstrated significant anti-infective, anti-inflammatory, antioxidant, pro-angiogenic, and pro-osteogenic effects in rat subcutaneous infection and femoral defect models.
Instead of relying on huge and expensive data centers for rolling out cloudbased services to rural and remote areas, we propose a hardware platform based on small single-board computers. The role of these micro-data centers is twofold. On the one hand, they act as intermediaries between cloud services and clients, improving availability in the case of network or power outages. On the other hand, they run community-based services on local infrastructure. We illustrate how to build such a system without incurring high costs, high power consumption, or single points of failure. Additionally, we opt for a system that is extendable and scalable as well as easy to deploy, relying on an open design.
P. Abrahamsson, S. Helmer, Tosin Daniel Oyetoyan et al.· arXiv.org· 2 citations
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduAug 27, 2026
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.