Self-Assembling Teams

• Self-assembling teams are groups formed through participant choice under defined constraints, ensuring bounded autonomy and clear roles.
• They are studied in agile software, educational projects, and online platforms, emphasizing dimensions like autonomy, shared leadership, and continuous learning.
• Empirical findings reveal that self-assembly enhances collaboration, balances workloads, and drives innovative task allocation through iterative self-assignment.

Searching arXiv for recent and foundational papers on self-assembling teams, self-organizing teams, and related team-formation frameworks. Self-assembling teams are teams whose composition, task allocation, or coordination pattern emerges primarily from participant choice rather than from detailed external assignment. Across the literature, the term appears in at least three closely related senses: students choosing their own teammates under instructor-defined constraints; agile software teams coordinating work through self-assignment and shared leadership; and online or AI-mediated systems in which people iteratively express teammate preferences while an algorithm reconciles those preferences with feasibility constraints. In all three senses, self-assembly is not treated as the absence of structure. It is treated as a bounded form of autonomy supported by explicit roles, artifacts, routines, and oversight mechanisms (Oza et al., 2013, Pinho et al., 3 Oct 2025, Lykourentzou et al., 2021).

1. Definitions and conceptual scope

In agile software research, self-organization is framed as a set of team-level capabilities rather than as leaderless improvisation. One influential lens, adopted in a student software project case study, characterizes self-organized teams through six interrelated dimensions: autonomy, team orientation, shared leadership, redundancy, learning, and communication and collaboration. Under this view, management “guides with mission” rather than detailed instructions, leadership shifts with task and expertise, and teams coordinate through shared artifacts and feedback loops rather than command-and-control (Oza et al., 2013).

In agile pedagogy, the pattern is formulated more explicitly as team formation: “Allow teams to self-assemble to create a sense of ownership and freedom of choice.” The immediate problem is how to assemble teams quickly while minimizing “initial social and inter-personal hurdles.” Self-assembly is therefore treated as a practical design decision for the setup phase of collaborative project courses, not merely as an emergent cultural phenomenon (Pinho et al., 3 Oct 2025).

In online work, self-assembly is defined even more sharply against centralized assignment. “Self-Organizing Teams (SOTs)” are described as “supported but not guided by an algorithm,” with workers themselves controlling teammate choice and collectively guiding the output of collaboration. The algorithm implements preferences at scale but does not impose a designer-specified notion of the best team (Lykourentzou et al., 2021).

A recurring misconception is that self-assembling teams are unstructured or anti-managerial. The literature points in the opposite direction. Research on autonomous agile teams explicitly treats “autonomous teams,” “self-organizing teams,” “self-managing teams,” and “empowered teams” as synonymous, but ties them to minimum critical specification, requisite variety, redundancy of function, and “learning to learn,” not to the removal of boundaries or governance (Stray et al., 2018).

2. Organizational properties and enabling conditions

The six dimensions used in agile software research provide a compact account of what a self-assembling team must be able to do. Autonomy means authority over work and participation in iteration planning and goal setting. Team orientation means shared responsibility and cross-functional composition. Shared leadership means a “lead-and-collaborate” principle in which leadership shifts with expertise. Redundancy means overlap in skills and conventions so members can cover for one another. Learning depends on continuous feedback, automated testing, continuous integration, and short review cycles. Communication and collaboration depend on close customer contact, daily information sharing, visible progress tracking, and, in the case study, working together in an open workspace (Oza et al., 2013).

These dimensions align closely with broader organizational principles summarized in research on autonomous agile teams. Minimum critical specification limits central prescription to goals and constraints. Requisite variety requires that team composition match environmental complexity. Redundancy of function makes replacement and assistance possible within the team. “Learning to learn” links autonomy to double-loop learning and continuous improvement. The same literature emphasizes autonomy, cross-fertilization, and self-transcendence as core conditions for genuinely autonomous teams (Stray et al., 2018).

The enabling conditions reported in empirical studies are consistent. Self-organization in the Helsinki Software Factory case was supported by a clear mission, a real customer, co-location, lightweight agile and lean process scaffolding, and some form of agile knowledge in the team or by a coach (Oza et al., 2013). In educational settings, self-assembly is enabled by explicit size bounds, a deadline for team registration, and a vetting process after which staff review and, if necessary, adjust team compositions (Pinho et al., 3 Oct 2025). In practice, the model is therefore autonomy within a designed envelope.

3. Mechanisms of assembly and reassembly

The literature distinguishes team formation from the everyday mechanisms through which teams continually reassemble around work. In university courses, the formation mechanism is straightforward: students form their own teams within instructor-defined bounds, typically in ranges such as $3 \leq n \leq 6$ depending on the course, then register the member list and often a contact person, technology profile, or topic preferences. Staff review the result, fix oversize or undersize teams, and place unassigned students into existing or new teams (Pinho et al., 3 Oct 2025).

In agile software development, the corresponding everyday mechanism is self-assignment. A grounded-theory study with 42 participants from 25 companies groups the drivers of self-assignment into task-based, developer-based, and opinion-based factors. Task learning potential is reported as the single most reported factor by developers, but choices are also shaped by business priority, technical complexity, dependencies, completion time, understandability, and desirability. Previous experience, technical expertise, workload, deference to co-workers, and preference for particular collaborators affect who clusters on which work. Managers influence sustainability through distinct styles characterized as risk-averse, risk-balancing, and growth-seeking (Masood et al., 2021).

In online work, self-assembly is operationalized as repeated preference expression. Workers in dyads complete a round of collaborative writing, evaluate one another, decide whether to stay with the current teammate, and nominate up to two alternatives. Preferences are encoded as directed scores, converted into undirected affinities, and matched by a greedy maximum-weight procedure over feasible dyads. The system therefore implements a mediated form of self-assembly: participants choose, the algorithm reconciles (Lykourentzou et al., 2021).

This body of work suggests that self-assembling teams are not formed once. They are continually re-formed through task pulls, teammate choices, and bounded reconfiguration. A plausible implication is that “team formation” and “task allocation” are analytically distinct but operationally entangled.

4. Performance and empirical evidence

The most direct qualitative link between self-organization and team performance comes from the Helsinki Software Factory case. The project team had 10 members, used a hybrid of XP, Scrum, and Lean practices, and developed an alpha-version prototype for a real customer. The study maps specific practices onto the six self-organization dimensions: backlog to autonomy; pair programming to team orientation; demo culture to shared leadership; co-located teams to redundancy; retrospectives and Kanban board to learning; and daily stand-ups, sprint planning, and Kanban board to communication and collaboration (Oza et al., 2013).

Performance in that study is not measured through velocity or defect counts. Instead, it is inferred through Hancock’s qualitative characteristics of high-performing teams. The strongest empirical associations are with autonomy and shared leadership, followed by communication and collaboration and learning. Redundancy showed no connection to the observed high-performance characteristics in that dataset, and “a belief in shared aims and objectives” showed no observed connection to any of the six self-organization dimensions in that case (Oza et al., 2013). The article’s practical import is narrow but clear: self-organizing behavior was most visible when the team made decisions by consensus, resolved conflicts themselves, and sustained free-flowing information.

Online crowd work provides a complementary quantitative result. In a controlled three-round collaborative writing task, teams in the SOT condition outperformed both a placebo condition and a no-agency condition on grammar, interest, originality, plot structure, and overall impression. For overall impression, mean ratings were $6.29$ in SOT versus $4.68$ and $4.70$ in the two benchmarks, with $F(2,1957)=99.847, p<.001, \eta^2=0.103$. The SOT condition also produced higher ratings of collaborator skillfulness, helpfulness, and collaboration ability, and more balanced turn-taking (Lykourentzou et al., 2021).

Large-scale observational evidence from GitHub shifts the focus from controlled interventions to emergent online project structure. A study of $N = 151{,}542$ repositories finds that while larger teams tend to be more successful, workload is highly concentrated, and highly successful teams are more focused than average teams of the same size. They also have more diverse experiential backgrounds and more members who are “leads” of other teams. In the multivariate model, the number of leads has the largest standardized coefficient, $\beta_L = 0.1388$, while experiential diversity also remains positive and significant (Klug et al., 2014). This does not use the language of self-assembly in pedagogy or agile process, but it provides a large-scale picture of bottom-up team growth and role concentration.

5. Scale, distribution, and multi-team coordination

In globally distributed agile teams, self-assembly becomes more visible in communication traces because coordination is mediated through shared tools. A psycholinguistic study of IBM Rational Jazz finds that practitioners enact different functional roles through language, and that team leaders were “most critical” to the teams’ self-organisation. Team leads showed the strongest collective orientation and broad role coverage; programmers were central contributors to information sharing, discussion, scaffolding, and comments; and enacted roles varied with the “cohort of features” being worked on (Licorish et al., 2021).

A related contextual-analysis study of the same environment emphasizes that these teams were extremely task focused and that team leads and programmers were central to self-organisation. The dominant interaction behaviors were information sharing, discussion, scaffolding, and comments. Formal roles existed, but enacted roles were more fluid: programmers routinely performed coordination and mentoring behaviors not implied by their formal title (Licorish et al., 2021). Self-assembling teams in distributed settings are therefore not role-free; they are characterized by role fluidity atop formal structure.

At program scale, coordination no longer consists only of self-managing feature teams. A four-year large-scale agile program with 12 development teams and about 175 people used a large ecology of mechanisms: program-level demos, Metascrum, Scrum of Scrums, project-level meetings in Business, Architecture, and Test, experience forums, Open Space Technology sessions, team boards, Jira, a wiki, an open work area, and “management by walking around.” A central finding is the “gradual transition to unscheduled meetings” as people got to know one another, allowing later coordination through direct informal interaction (Dingsøyr et al., 2018). This suggests that large-scale self-assembly depends on a scaffolding phase in which relationships and transactive memory are deliberately built before more decentralized patterns can dominate.

Research on autonomous agile teams reinforces the same boundary condition from a more conceptual angle. Too many dependencies, weak coaching, lack of trust, unclear goals, diversity in norms, and conflicts around roles all undermine autonomy. The literature also notes that autonomy is beneficial when task interdependence is high and can have a negative effect when interdependence is low (Stray et al., 2018). Self-assembling teams are therefore not universally superior; they are highly context-sensitive.

6. Computational and AI-mediated formulations

A substantial branch of the literature translates self-assembling teams into optimization and learning problems. In the Synergistic Team Composition Problem, the goal is to find a size-constrained partition $P_m$ that maximizes a product of team synergies,

$$P^*_m = \arg\max_{P_m\in\mathcal{P}_m(A)} S(P_m,\tau),$$

where team synergy combines proficiency and congeniality. Proficiency is defined through competence coverage and penalties for under- and over-proficiency; congeniality incorporates diversity on selected personality axes, the presence of specific personality profiles, and gender balance. The Bernoulli–Nash product favors balanced partitions rather than a few exceptionally strong teams (Andrejczuk et al., 2019).

Conflict-aware team formation generalizes the centralized problem further by combining capacities, task preferences, and pairwise conflicts in a conflict graph. The objective is to maximize

$\sum_{i \in I} \sum_{t \in T} u_{i,t} x_{i,t}$

subject to individual assignment, task capacity, and conflict constraints such as

$6.29$0

Approximation algorithms based on dependent rounding are proposed for this setting (Nikolaou et al., 2024). This is not self-assembly in the strong human-agency sense of online SOTs, but it formalizes a recurrent tension in the literature: how to reconcile preferences with global feasibility and incompatibility constraints.

AI-augmented formulations make that reconciliation sequential and adaptive. One dissertation models team formation as a multi-armed bandit problem in which each candidate team is an arm and user feedback is the reward. It later solves a global assignment problem using learned preference scores. The same work introduces tAIfa, an LLM-based feedback assistant that uses metrics including Language Style Matching, sentiment, Transactive Memory System indicators, balanced participation, pronoun use, communication flow, and topic coherence to nudge teams during the performing stage, and PuppeteerLLM, an LLM-based simulation framework for exploring team dynamics under different policies (Almutairi, 5 Jun 2025). In this line of work, self-assembly becomes assistive rather than purely spontaneous: humans express preferences, the algorithm learns, and assignment remains globally coordinated.

7. Limitations and contested issues

The literature is unusually consistent in warning against romantic interpretations of self-assembly. In education, self-selection can create friendship-based teams that ignore skills and experience, produce resentment when lone students must be added to existing teams, and still leave room for conflict and dysfunction. Instructors therefore retain a “final say” through vetting and post-deadline adjustment (Pinho et al., 3 Oct 2025).

In agile work allocation, self-assignment can become unsustainable when business priority is routinely overridden by learning potential or desirability, when the same experts repeatedly take critical work, when undesirable tasks are orphaned, or when quiet and deferential developers are systematically overlooked. Managerial support is therefore treated as necessary to sustain, not suppress, self-assignment (Masood et al., 2021).

The online-work literature identifies a different risk profile: popularity bias and exclusion. Because participants tend to prefer previous winners, self-assembling systems can create clusters of attractive and peripheral workers unless mitigations such as serendipity, diversity-aware recommendation, or mentorship incentives are introduced (Lykourentzou et al., 2021).

Empirical generalizability is also limited. The Helsinki case is a single student software project and explicitly a “preliminary empirical study” without quantitative productivity metrics (Oza et al., 2013). The Jazz studies are based on one product in one organization (Licorish et al., 2021, Licorish et al., 2021). The GitHub study is correlational and uses stargazers as a proxy for success (Klug et al., 2014). Research on autonomous agile teams explicitly states that there is “no one-size-fits-all autonomy approach” and highlights legal, architectural, contractual, and organizational constraints as persistent limits on self-organization (Stray et al., 2018).

Taken together, these limitations imply that self-assembling teams are best understood not as a universal organizational endpoint but as a family of bounded coordination arrangements. They depend on explicit constraints, visible work, mechanisms for feedback and conflict resolution, and, in many settings, lightweight but persistent forms of leadership and mediation.