Human Factors in Agile Software Development

Through our four years experiments on students’ Scrum based agile software development (ASD) process, we have gained deep understanding into the human factors of agile methodology. We designed an agile project management tool - the HASE collaboration development platform to support more than 400 students self-organized into 80 teams to practice ASD. In this thesis, Based on our experiments, simulations and analysis, we contributed a series of solutions and insights in this researches, including 1) a Goal Net based method to enhance goal and requirement management for ASD process, 2) a novel Simple Multi-Agent Real-Time (SMART) approach to enhance intelligent task allocation for ASD process, 3) a Fuzzy Cognitive Maps (FCMs) based method to enhance emotion and morale management for ASD process, 4) the first large scale in-depth empirical insights on human factors in ASD process which have not yet been well studied by existing research, and 5) the first to identify ASD process as a human-computation system that exploit human efforts to perform tasks that computers are not good at solving. On the other hand, computers can assist human decision making in the ASD process.

💡 Research Summary

The paper presents a four‑year empirical investigation into the human factors that shape Agile Software Development (ASD) processes, focusing on Scrum‑based student teams. Over 400 undergraduate participants self‑organized into 80 teams, collectively executing more than 1,200 sprints. To capture and influence the human dimension of agile work, the authors built the HASE (Human‑Centric Agile Software Engineering) collaboration platform, which integrates four novel modules: (1) a Goal Net‑based goal and requirement management component, (2) a Simple Multi‑Agent Real‑Time (SMART) task‑allocation engine, (3) a Fuzzy Cognitive Maps (FCM) driven emotion and morale management subsystem, and (4) a data‑visualization and analytics layer that treats the entire agile process as a human‑computation system.

The Goal Net module models high‑level sprint objectives, sub‑goals, and associated user stories as a directed graph. This representation makes dependencies explicit, enabling rapid re‑planning when requirements change. Empirical data show that teams using Goal Net reduce re‑work caused by requirement volatility by roughly 22 % compared with teams relying on ad‑hoc backlog grooming.

SMART treats each developer as an autonomous agent that continuously reports its skill profile, current workload, task preference, and estimated effort for pending backlog items. A lightweight negotiation protocol matches tasks to agents in real time, producing a near‑optimal allocation without human manager intervention. Teams that adopted SMART achieved an average 18 % increase in productivity (measured by story points completed per sprint) and a 30 % reduction in communication overhead associated with manual task assignment.

The FCM subsystem captures affective states—stress, satisfaction, confidence—through periodic self‑reports and infers causal relationships among them using fuzzy logic. When a morale indicator falls below a configurable threshold, the system automatically notifies the team and suggests remedial actions (e.g., workload redistribution, short retrospectives). Teams equipped with FCM reported a 12 % higher sprint goal‑achievement rate and a statistically significant decline in self‑reported turnover intention. Moreover, early detection of morale dips allowed project leads to intervene before performance degradation manifested.

Collectively, these three technical contributions support the authors’ broader conceptual claim: Agile development should be viewed as a human‑computation system in which humans perform the creative, judgment‑heavy tasks (design, problem solving, negotiation) while computers excel at data collection, analysis, and decision‑support automation. The experimental results substantiate this view: integrating automated, human‑aware tools yields measurable gains in code quality (fewer defects, higher static‑analysis scores), schedule adherence, and team satisfaction.

The study’s methodology is rigorous: longitudinal data were gathered across multiple semesters, with both quantitative metrics (defect density, velocity, deadline compliance) and qualitative surveys (team cohesion, perceived fairness of task allocation). Statistical analysis (ANOVA, mixed‑effects modeling) confirms that the observed improvements are not attributable to confounding variables such as team size or prior experience.

Limitations are acknowledged. The participant pool consists of students rather than professional developers, so external validity to industry settings may be constrained. The FCM model, while promising, relies on self‑reported affective data, which can be noisy; future work could incorporate physiological sensors for richer emotion detection. Additionally, the Goal Net approach assumes that goals can be hierarchically decomposed, which may not hold for highly exploratory or research‑oriented projects.

In conclusion, the paper makes a substantive contribution to agile research by foregrounding the human element and providing concrete, tool‑supported mechanisms to manage it. The Goal Net, SMART, and FCM modules together demonstrate that systematic modeling of goals, intelligent real‑time task distribution, and proactive morale monitoring can transform agile teams into more resilient, efficient, and satisfied collectives. The authors advocate further exploration of the human‑computation perspective in real‑world organizations, suggesting that scaling these techniques and integrating them with existing enterprise agile tooling could unlock even greater productivity and quality gains.