Alignment-Process-Outcome: Rethinking How AIs and Humans Collaborate

Alignment–Process–Outcome: Rethinking How AIs and Humans Collaborate Haichang Li hli52@gmu.edu George Mason University Fairfax, Virginia, USA Anjun Zhu aza99@sfu.ca Simon Fraser University Burnaby, British Columbia, Canada Arpit Narechania arpit@ust.hk The Hong Kong University of Science and Technology Hong Kong S.A.R., China Figure 1: Overview of our two-lens view of collaboration flow. Left: the task lens models collaboration as a trajectory evolving in a task space from the Problem Root toward the Solution Circle. Top-right: the intent lens shows how participants’ implicit intents become explicit and aggregate into a collective intent. Bottom-right: examples of potential trajectory updates at a local decision point in the task space (e.g., forward progression, backtracking, revisiting prior alternatives, and branching). Abstract In real-world collaboration, alignment, process structure, and out- come quality do not exhibit a simple linear or one-to-one correspon- dence: similar alignment may accompany either rapid convergence or extensive multi-branch exploration, and lead to different results. Existing accounts often isolate these dimensions or focus on specific participant types, limiting structural accounts of collaboration. We reconceptualize collaboration through two complementary lenses. The task lens models collaboration as trajectory evolution in a structured task space, revealing patterns such as advancement, branching, and backtracking. The intent lens examines how individ- ual intents are expressed within shared contexts and enter situated decisions. Together, these lenses clarify the structural relationships among alignment, decision-making, and trajectory structure. This work is licensed under a Creative Commons Attribution 4.0 International License. CHI EA ’26, Barcelona, Spain © 2026 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-2281-3/2026/04 https://doi.org/10.1145/3772363.3799032 Rather than reducing collaboration to outcome quality or treating alignment as the sole objective, we propose a unified dynamic view of the relationships among alignment, process, and outcome, and use it to re-examine collaboration structure across Human–Human, AI–AI, and Human–AI settings. CCS Concepts • Human-centered computing→Collaborative and social computing theory, concepts and paradigms. Keywords Collaboration Flow, Collaboration Dynamics, Alignment, Task Out- come ACM Reference Format: Haichang Li, Anjun Zhu, and Arpit Narechania. 2026. Alignment–Process– Outcome: Rethinking How AIs and Humans Collaborate. In Extended Ab- stracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26), April 13–17, 2026, Barcelona, Spain. ACM, New York, NY, USA, 7 pages. https://doi.org/10.1145/3772363.3799032 arXiv:2603.08017v2 [cs.HC] 10 Mar 2026 CHI EA ’26, April 13–17, 2026, Barcelona, SpainLi, Zhu, and Narechania 1 Introduction In real-world settings, alignment, collaboration process, and task outcome often exhibit a persistent disconnect rather than a stable monotonic positive relationship. This means that perfect align- ment does not inherently guarantee a smooth, linear process or a superior outcome, as these three dimensions are not naturally coupled. Alignment or misalignment can emerge in collaboration with multi-branch exploration or near-linear progress, and both can lead to strong or weak outcomes. For example, a team that initially struggles to understand each other may still reach a high-quality solution through iterative detours and exploration, while a highly aligned team may progress smoothly yet converge prematurely to a merely average solution. This disconnect reveals a critical gap: without a unified frame- work to understand the structural relationships among these three dimensions, research and practice tend to adopt fragmented per- spectives, creating problematic blind spots. For example, alignment- centric views risk mistaking consensus for quality, overlooking pre- mature convergence to mediocre solutions. Process-centric views may penalize exploratory backtracking as inefficiency, failing to dis- tinguish productive iteration from genuine waste. Outcome-centric views can obscure whether intent participation was equitable or whether collaboration was merely nominal. Meanwhile, collaboration research has expanded from traditional Human–Human interaction to AI–AI and Human–AI settings. Yet the underlying theories remain fragmented by participant type and scope of analysis, contributing to these siloed perspectives. Collab- oration is often discussed separately across different participant configurations [16,39,44,47]. Macro-level work often characterizes collaboration through stage models or task categories [12,28,40], whereas micro-level work focuses on local interaction mechanisms and behavioral details, such as grounding and conversational repair [10,34]. These lines of work are valuable on their own. However, they lack a shared structural discussion to explain how individual intents aggregate and diverge during collaboration, and how they gradually shape the macro-level trajectory of task progress. Accordingly, we argue for the need to rethink how humans and AIs collaborate. Some classic theories have also been explicitly framed as descriptive abstractions that generalize empirical pat- terns across different work domains [35]. Rather than following an empirical path that induces patterns from records of completed collaborations and treats the process as a static object for retrospec- tive analysis, we call for a complementary foundational perspective that derives the evolution of collaboration from the underlying dynamics of the process itself. This perspective starts from task structure and intent aggregation, conceptualizing collaboration as a continuously evolving dynamic process—termed collaboration flow. By examining the dynamics and trajectories of collaboration flow, we re-examine the structural relationships among alignment, col- laboration, and task outcomes, providing a more generative starting point for future research and system design. 2 Two Lenses for Rethinking Collaboration We introduce two complementary lenses for observing collabora- tion: task lens and intent lens. The task lens focuses on the overall evolution from start to finish within the task structure. The intent lens moves from individual to collective, examining how intents are expressed, perceived, and aggregated in collaboration. Together, these lenses analyze collaboration flow and support subsequent discussion of the structural relationships among alignment, collab- oration, and task outcomes. 2.1 Task Lens: From Start to Finish On the macro level, the task lens treats collaboration as a holistic process, examining its overall evolutionary trajectory from start to finish within the task structure. From the task structure perspective, collaboration often begins with a well-defined problem, yet potential solutions and evolution- ary paths can take many different forms, making them difficult to characterize by a single dimension. To describe this overall evolu- tion, we abstract the task structure as a task space that captures collaboration from start to finish: a 2D-Disk centered at the Problem Root, with its boundary defined as the Solution Circle. This disk abstraction stems from a natural deduction of the task’s expansive nature: since all collaborative efforts originate from a single root and radiate toward infinite potential outcomes, the task space is inherently radiating in nature. Furthermore, because every final solution represents a 100% completion state regardless of its spe- cific form, these endpoints are logically equidistant from the root in terms of progress, naturally constituting a circular boundary—the solution circle. Collaboration is thus characterized as an evolutionary trajec- tory from the problem root toward the solution circle. Following a polar-coordinate intuition, the radial direction represents collabo- ration depth, indicating the degree to which the collaboration state advances from the problem root (푑=0%) toward the solution cir- cle (푑=100%). In this coordinate system, the radial progression directly reflects the reduction in distance to the final goal and solu- tion. Radial progression need not increase monotonically over time; collaboration may stall or involve backtracking, revisiting earlier branches and partially undoing progress. The angular direction corresponds to path variation, capturing path choices that lead to differences in final solution forms at similar collaboration depth. This reflects exploration across different branches of the task space. We illustrate this 2D-Disk presentation in Fig. 1 to provide a shared coordinate system, where the radial direction represents col- laboration depth and the angular direction captures path variation. The concentric circles illustrated in the figure serve as conceptual markers provided solely for abstraction and clarity rather than fixed temporal milestones. This distinction accounts for the fact that, in practice, the actual increments of progress resulting from collab- orative decisions are often irregular and non-linear, representing discrete movements across different levels of progression within this continuous radial space. We also use this representation to support the interface design of our future trajectory visualization work in Section 4.3. Under this lens, task completion is determined by radial position, while the final solution is determined by angular path choices; the intersection of the trajectory with the solution circle corresponds to the final solution. Trajectory features (such as backtracking frequency, angular deviation, and branching count) provide observ- ables for discussing collaboration efficiency and exploration. Alignment–Process–Outcome: Rethinking How AIs and Humans CollaborateCHI EA ’26, April 13–17, 2026, Barcelona, Spain 2.2 Intent Lens: From Individuals to Collective While the trajectory patterns in the task lens (e.g., backtracking frequency or angular shifts) macroscopically reflect collaborative efficiency and exploration, this perspective obscures the composi- tion of local states. Specifically, it masks implicit and conflicting individual goals, as well as potential paths that remain unselected. To address this, we introduce the intent lens as a complementary micro-level analysis to examine how the collaboration flow is con- structed from individual to collective intents. Whether human or AI, participants possess both directly ex- pressed Explicit Intents and Implicit Intents that cannot be directly perceived. The latter includes unexpressed goals, preferences, and tacit knowledge—internal constraints that continuously influence judgment and decision-making. To function in collaboration, im- plicit intents must be projected into the shared context through multimodal semantic channels such as language and visualization, becoming explicit intents that others can perceive, understand, and respond to. This process is context-dependent: the representation of intentions in semantic space depends not only on their expressed content but also on the surrounding context. For example, “I’m okay with that” may be interpreted as explicit support or reluctant compliance depending on the collaborative situation. Consequently, relative relationships between intentions are often difficult to com- pare directly across contexts. Conceptually, intents can be represented in a computational form. These semantic intents (e.g., natural language, gestures, expres- sions) can be represented via embeddings and compared through similarity metrics [9,41]. However, similarity alone is insufficient to characterize intent relationships in collaboration. For instance, “I support Plan A” and “I oppose Plan A” may be semantically similar yet exert opposite effects on collaboration progress [20]. Therefore, representing how individual intents aggregate into collective deci- sions requires encoding the relative stance between intents beyond similarity [19], reflecting differences in progress direction. Collaboration can thus be viewed as a series of decision points triggered by intent relationships within local task contexts. At these nodes, intentions are perceived, compared, and weighed, jointly determining whether the collaboration flow advances, branches, or backtracks. The accumulation of local decisions forms the overall trajectory of the flow. The task lens takes the collective as the unit of analysis, character- izing the overall collaboration trajectory within the task structure. The intent lens focuses on how individual intentions aggregate into collective decisions, thereby driving the flow of collaboration. Neither lens imposes restrictions on entities, together forming an analytical language for understanding the nature of collaboration across Human–Human, AI–AI, and Human–AI scenarios. 3 Theoretical Deductions and Preliminary Findings This section summarizes a set of preliminary understandings regard- ing the relationships among alignment, the collaboration process, and task outcomes. These findings are derived from a natural deduction of our framework, where the intent lens defines the pre- requisites for interaction and the task lens provides the geometric dimensions of the collaboration space. The logic of our derivation is straightforward: since the intent lens requires a “shared context” for sensemaking and the task lens defines “radial” and “angular” directions, alignment naturally emerges as a three-level relation (Contextual, Radial, and Angular). Furthermore, because individual intents aggregate into a “collective intent” through a weighting process—where ignoring a viewpoint is functionally equivalent to assigning it a weight of zero—the trajec- tory’s movement becomes a direct manifestation of this weighted decision-making. Finally, because the radial and angular directions in a polar coordinate system are orthogonal and independent, we deduce a structural decoupling between process and outcome. Ulti- mately, this framework allows us to re-characterize the fundamental interplay between alignment (the modulator), collaboration flow (the execution), and outcome (the destination). 3.1 Alignment as a Multi-Level Relation Recent work suggests that framing alignment as a binary state of “whether aligned” is insufficient [18,23,36,38]. Through the joint application of the two lenses, alignment emerges as a multi-layered relationship operating at different levels. Building on the introduced radial–angular task lens, and the context-dependence of intent sensemaking, we note three levels of alignment: Contextual, Radial, and Angular, characterizing their distinct roles in collaboration. Contextual Alignment characterizes whether participants’ intents reside within the same shared context and can be correctly under- stood by one another. When contextual alignment is insufficient, misunderstanding blocks effective interaction even if response mo- tivation or stance association exists, preventing collaboration. Building on this foundation, Radial Alignment governs whether collaboration can continuously advance toward task completion. When participants diverge on the direction of radial progress, collab- oration flow may stall or backtrack, reducing or even resetting col- laboration depth. Even when intents are semantically highly related, negative stance can offset radial progress and trigger depth regres- sion, reintroducing previously compressed angular path choices. When both contextual and radial alignment hold, the focus of alignment shifts to Angular Alignment, which characterizes the relative relationships between intents in path selection, shaping the structural form of collaboration flow. When intents are highly related with aligned stance, collaboration tends to advance radially. When intents are related but stance diverges, angular tension is more likely to translate into branching and parallel exploration rather than radial backtracking. For intents with weak or ambigu- ous stance, highly related intents typically enrich existing paths through information supplementation, while weakly related intents are more likely to introduce new focal points (e.g., shifting from technical implementation to user experience goals). From this perspective, contextual alignment determines if collab- oration can begin, radial alignment affects if it can sustain progress, and angular alignment shapes the path structure in the task space. This layered view explains why similar alignment levels can yield different collaboration trajectory shapes and outcomes. CHI EA ’26, April 13–17, 2026, Barcelona, SpainLi, Zhu, and Narechania 3.2 Collaboration as a Weighted Decision-Making Process Whether collaboration sustains progress depends on how multiple intents are weighted and aggregated at critical decision points, and alignment primarily modulates the cost and stability of this weight- ing process. From the intent lens, advancement, backtracking, and branching can be understood as decision outcomes produced by intent weighting. Weight allocation mechanisms determine which intents enter decision-making and their degree of influence, shap- ing the dynamics and path structure of collaboration. Considering common organizational constraints, we preliminarily distinguish between structured allocation and negotiated allocation, correspond- ing to different roles of alignment. In practice, weighting does not occur in a vacuum. In struc- tured weight allocation, weight generation is strongly constrained by established agency or protocols (e.g., high-agency individuals, voting, or rotation rules). Collaboration may thus advance despite insufficient alignment, though it risks degenerating into nominal collaboration—unless low-weight intents substantially expand or reshape the collective decision space. In contrast, negotiated weight allocation leaves weights unlocked; participants retain decision con- tribution rights and spontaneously form weighting schemes. Align- ment becomes critical for stable weight convergence. Misalignment prevents convergence, forcing teams to invest extra costs in estab- lishing shared context and reaching consensus on depth and path choices. Collaboration flow becomes prone to costly oscillations or deadlock. Once weights form, multiple intents collapse into a single collec- tive intent that updates the trajectory; different weighting regimes can therefore yield different exploration patterns and land on dif- ferent regions of the solution circle. 3.3 Rethinking the Relationship Between Alignment, Collaboration Process, and Outcome As previously discussed, collaboration process and task outcome are structurally decoupled: outcome depends on which path variations are retained or compressed during collaboration, not on whether the trajectory is tortuous or linear. The key lies in how weight allocation aggregates multiple intents into collective intent and determines branch retention at critical nodes, causing collaboration to land in different regions of the solution circle. Section 3.1 further discusses how misalignment exerts compounded effects on this mechanism by altering weight convergence and path retention. Building on this, we examine cases where all three alignment levels hold: collaboration often exhibits rapid and nearly linear radial progression. However, high efficiency does not necessarily guarantee high-quality outcomes. Since angular divergence may be systematically compressed, collaboration flow can still prematurely converge to local optima or structurally biased regions. Conversely, moderate intent divergence, while increasing coordination costs, may introduce path tension and trigger weight reallocation, en- abling the collaboration trajectory to cover a broader solution space. Based on this preliminary understanding, we view alignment as a variable modulating collaboration cost and stability, weight allo- cation as the execution mechanism driving collaboration forward, and outcome as the landing point of the collaboration trajectory in the solution space. Completing a collaboration merely means the trajectory reaches the solution circle, without guaranteeing optimality of its endpoint in task terms. 4 Proposed Calls and Implications Building on the above reasoning, we propose the following calls and describe our work-in-progress system. 4.1 Revisiting Metrics: Trajectory, Efficiency, and Outcome Collaboration quality cannot be adequately proxied by any sin- gle signal. Focusing solely on outcome systematically flattens pro- cess structure, such as exploration scope, convergence patterns, and whether intents enter critical decisions. Focusing solely on alignment risks misinterpreting low-friction progression as high- quality collaboration, overlooking outcome quality and path cover- age. Optimizing only efficiency or exploration may ignore intent relationships and weight aggregation mechanisms, yielding favor- able process metrics but endpoints that deviate from goals. We thus call for joint metrics characterizing collaboration by simulta- neously considering outcome quality, progression efficiency, and trajectory structural features (advancement, branching, backtrack- ing). Meanwhile, if systems treat backtracking as failure, weight allocation will systematically penalize exploratory intents. Instead, backtracking should be viewed as a low-cost, recordable, reusable collaboration product. Quantifying how intents enter decisions requires collaboration-oriented intent representation: encoding stance direction relative to shared goals beyond semantic similarity, distinguishing semantically similar intents with opposite direc- tional effects, and linking their impact on weight aggregation and trajectory evolution. 4.2 Rethinking System Strategies: Structural Plasticity and Strategic Friction Weights should not be rigidly bound to collaborator identity. We call for treating weight allocation as a context-switchable protocol layer, reversibly transitioning between negotiated and structured allocation: strengthening structural constraints during oscillations to restore progression, and preserving controlled divergence and parallel exploration when convergence risks arise, maintaining path diversity. Under the coupling of alignment and weight allocation, premature convergence is a common structural risk. Early-stage collaboration should thus moderately preserve divergence rather than uniformly suppress it. In certain contexts, introducing moder- ate friction (e.g., reflection roles or delayed critical confirmations) may prove more effective [7,8,15]. This further points to an AI design strategy: AI agents should not blindly defer to humans but introduce interpretable strategic friction at critical nodes to miti- gate bias and premature convergence risks. As ongoing work, we are exploring agent support strategies for different misalignment scenarios and integrating them with the trajectory visualization interface discussed later. Alignment–Process–Outcome: Rethinking How AIs and Humans CollaborateCHI EA ’26, April 13–17, 2026, Barcelona, Spain 4.3 Redesigning Interface Paradigms: From Linear Streams to Topological Navigation Collaboration can be understood as trajectory evolution in task space; linear collaboration logs obscure branching, backtracking, and unchosen paths, impeding review, comparison, and diagno- sis. We call for shifting collaboration interfaces from linear flows to navigable trajectory and topological representations, enabling users to explicitly examine progression depth, critical branches and backtracking, and understand how intents aggregate into actions at decision nodes [11,13]. As ongoing work, we explore visual- ization and interaction mechanisms for collaboration trajectories and examine their embedding in real workflows. For example, in conversational AI IDEs such as Cursor [5], task phases, critical decision points, and historical exploration paths could be presented as navigable structures, supporting comparison, backtracking, and reuse. At the same time, this interface design is generalizable to a wider range of application contexts, including writing and design. 5 Discussion and Future Work Although the above analysis avoids binding collaboration mecha- nisms to any specific participant type, we build on these preliminary results to discuss how they manifest differently across collaboration settings and to clarify the structural factors that require additional attention when rethinking how humans and AIs collaborate. 5.1 Discussions Under Different Participant Type Settings In Human–Human collaboration, the weight allocation process is often pre-shaped by hierarchy and social norms [1,42], such that apparent progress may stem from preset weights rather than meaningful intent participation. For instance, passive deference to a superior person can compress path exploration and induce nominal collaboration and premature convergence. DAOs [14] can be seen as attempts to decouple weights from identity; however, without stable constraints, negotiated weight allocation can easily turn into high-friction oscillation. Together with large individual capability gaps and substantial intent projection loss [31], designs that reduce the cost of contextual alignment (e.g., slides [26] or explicit terminology explanations [6]) often become essential for stabilizing the collaboration flow. In AI–AI collaboration, performance differences across agents are typically less diverse. Since social hierarchies do not exist in this case, there are no social anchors to limit the weight allocation pro- cess, making weights more likely to degenerate into high-friction negotiation. Thus collaboration flow may constantly oscillate at decision points instead of reaching a stable conclusion, which ex- plains why multi-agent systems are not necessarily better than single-model execution; recent evidence suggests that as base mod- els become more capable, the marginal gains from multi-agent systems can shrink and may even reverse [25]. Systems therefore often introduce structured weights (e.g., manager agents [43], work- flow orchestration [30], or protocols [4]) to restore progress, and rely on cross-task memory [48], explicit reasoning [29], and context engineering [3] to reduce intent projection loss. Yet overly strong structure can also systematically compress the path space, allowing faster convergence to the solution circle while increasing the risk of landing in biased regions. In Human–AI collaboration, the weight boundaries are usually set by humans, but in practice they can also be dynamically ad- justable: local decisions can be delegated to AI through explicit authorization (e.g., Full Permission in coding agent settings [2,5]); while humans can actively reclaim control when the stakes escalate or when novel, high-risk, or nuanced situations arise [22]. Vibe Coding [24,33] suggests that even under imperfect alignment be- tween human intent and AI execution, structured weights can still sustain progress and yield usable outcomes. Meanwhile, the intent loss during the translation phase from human brain to human-side Human-AI interface is inevitable. For instance, previous work finds that people often fail to construct and iteratively improve effec- tive prompts for coding [45], and the resulted cognitive bias can lead to sub-optimal outcomes [50]. This unavoidable intention loss means that certain methods including carefully designed prompt engineering [32], intent-augmentation [17,27], and low-barrier AI interaction [49], are necessary to help shape the collaboration tra- jectory. However, these interventions often work by making weight allocation and intent expression more controllable and inspectable. The central challenge, therefore, is to support interpretable weight switching [21,37,46]–a long-standing HCI debate on user control, agent autonomy, and mixed-initiative interaction. Yet, preventing interfaces from obscuring the collaboration trajectory and AI’s in- termediate work at the same time else it is difficult to inspect how and why collaboration converged to a particular solution region. 5.2 Future Work and Downstream Effects Our discussion is primarily conceptual and qualitative. The geo- metric formulation serves as a structural metaphor rather than a formal characterization. Operationalizing this framework requires further work along multiple directions. To address this limitation, future efforts should transform traditional linear collaboration logs into a topological navigation interface that explicitly visualizes collaboration flow. Such an interface would enable users to inspect and reflect on the collaboration process. To support quantitative analysis, it is necessary to develop intent vector encoding methods that go beyond simple semantic similarity. This includes construct- ing a computational approach that jointly represents intent content, similarity, and stance relations. We refer to this representation as intent embedding. In terms of downstream effects, this framework offers a novel perspective for the HCI community to re-examine how we observe, diagnose, and design collaboration. At the observation level, it shifts attention from isolated, surface-level phenomena to insights into structural patterns. For example, when a team reaches consensus quickly, the framework invites us to use angular coverage of the trajectory to assess whether this reflects efficient progress or prema- ture convergence driven by insufficient exploration of alternatives. At the diagnosis level, the framework turns ambiguous explana- tions into traceable mechanisms. By characterizing the morphology of collaborative trajectories and conducting statistical analyses at key decision points, we can intuitively describe collaborative per- formance. For instance, by examining how branching and revisit CHI EA ’26, April 13–17, 2026, Barcelona, SpainLi, Zhu, and Narechania behaviors relate to alignment at different levels, we can offer ac- tionable recommendations and precisely articulate the health of collaboration. At the design level, the framework points to a shift from single-objective optimization to structured balancing. Design- ers can preserve angular divergence early to avoid premature con- vergence. They can introduce strategic friction at critical junctures to sustain path diversity. They can also use switchable weighting protocols to ensure that radial progress does not stagnate. More broadly, this perspective can ground a foundational framework for more comprehensive reasoning about collaboration. It encourages the community to re-examine the structure of collaboration, rather than explaining collaboration solely through outcomes or treating alignment—especially in human–AI settings—as the only objective. This structural approach provides researchers with a more princi- pled empirical basis, system designers with theoretically grounded design guidelines, and practitioners with higher-dimensional aware- ness and control over collaborative processes. 6 Conclusion We take the position that the relationship among alignment, collab- oration process, and task outcome is not a stable, one-dimensional positive correlation. We introduce two analytic lenses—task and intent—to view collaboration as trajectory evolution within a task, driven by how multiple intents are weighted and aggregated into decisions. Under this view, alignment modulates coordination cost, while weight allocation determines whether exploration is pre- served or compressed, shaping both trajectory structure and out- come quality. We hope this perspective can inform future metrics, system strategies, and interface paradigms for collaboration be- tween and among humans and AIs. References [1]Wataru Akahori, Naomi Yamashita, Jack Jamieson, Momoko Nakatani, Ryo Hashimoto, and Masahiro Watanabe. 2023.Impacts of the Strength and Conformity of Social Norms on Well-Being: A Mixed-Method Study Among Hybrid Workers in Japan. In Proceedings of the 2023 CHI Conference on Hu- man Factors in Computing Systems (Hamburg, Germany) (CHI ’23). Associa- tion for Computing Machinery, New York, NY, USA, Article 870, 17 pages. doi:10.1145/3544548.3581383 [2]Anthropic. 2025. Claude Code: AI coding agent for terminal and IDE. https: //w.claude.com/product/claude-code. Accessed: 2026-01-21. [3]Anthropic. 2025. Effective Context Engineering for AI Agents. https://w. anthropic.com/engineering/effective-context-engineering-for-ai-agents. Pub- lished 29 September 2025; Accessed: 2026-01-21. [4]Anthropic. 2025.Equipping Agents for the Real World with Agent Skills. https://w.anthropic.com/engineering/equipping-agents-for-the-real- world-with-agent-skills. Published 16 October 2025; Accessed: 2026-01-21. [5] Anysphere. 2026. Cursor. https://w.cursor.com/. Accessed: 2026-01-22. [6]Anastasia Axaridou, Konstantina Konsolaki, Maria Theodoridou, Artem Kozlov, Peter Haase, and Martin Doerr. 2018. VisTA: Visual Terminology Alignment Tool for Factual Knowledge Aggregation. In Third International Workshop on Semantic Web for Cultural Heritage (SW4CH 2018). [7]Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z. Gajos. 2021. To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI- assisted Decision-making. Proc. ACM Hum.-Comput. Interact. 5, CSCW1, Article 188 (April 2021), 21 pages. doi:10.1145/3449287 [8] Zeya Chen and Ruth Schmidt. 2024. Exploring a Behavioral Model of “Positive Friction” in Human-AI Interaction. In Design, User Experience, and Usability: 13th International Conference, DUXU 2024, Held as Part of the 26th HCI International Conference, HCII 2024, Washington, DC, USA, June 29–July 4, 2024, Proceedings, Part I (Washington DC, USA). Springer-Verlag, Berlin, Heidelberg, 3–22. doi:10. 1007/978-3-031-61353-1_1 [9]Qingrong Cheng, Xu Li, and Xinghui Fu. 2024. SIGGesture: Generalized Co- Speech Gesture Synthesis via Semantic Injection with Large-Scale Pre-Training Diffusion Models. In SIGGRAPH Asia 2024 Conference Papers (Tokyo, Japan) (SA ’24). Association for Computing Machinery, New York, NY, USA, Article 133, 11 pages. doi:10.1145/3680528.3687677 [10]Herbert H Clark and Susan E Brennan. 1991. Grounding in communication. (1991). [11]Adam J Coscia, Shunan Guo, Eunyee Koh, and Alex Endert. 2025. OnGoal: Tracking and Visualizing Conversational Goals in Multi-Turn Dialogue with Large Language Models. In Proceedings of the 38th Annual ACM Symposium on User Interface Software and Technology (UIST ’25). Association for Computing Machinery, New York, NY, USA, Article 208, 18 pages. doi:10.1145/3746059. 3747746 [12]Design Council. 2005. The Double Diamond. https://w.designcouncil.org.uk/ our-resources/the-double-diamond/ [13] Zijian Ding, Michelle Brachman, Joel Chan, and Werner Geyer. 2025. "The Diagram is like Guardrails": Structuring GenAI-assisted Hypotheses Exploration with an Interactive Shared Representation. In Proceedings of the 2025 Conference on Creativity and Cognition (C&C ’25). Association for Computing Machinery, New York, NY, USA, 606–625. doi:10.1145/3698061.3726935 [14] Youssef El Faqir, Javier Arroyo, and Samer Hassan. 2020. An overview of decen- tralized autonomous organizations on the blockchain. In Proceedings of the 16th International Symposium on Open Collaboration (Virtual conference, Spain) (Open- Sym ’20). Association for Computing Machinery, New York, NY, USA, Article 11, 8 pages. doi:10.1145/3412569.3412579 [15]Jonathan Ericson. 2023. Reimagining the Role of Friction in Experience Design. J. User Exper. 17, 4 (June 2023), 131–139. [16]George Fragiadakis, Christos Diou, George Kousiouris, and Mara Nikolaidou. 2025. Evaluating Human-AI Collaboration: A Review and Methodological Frame- work. arXiv:2407.19098 [cs.HC] https://arxiv.org/abs/2407.19098 [17] Frederic Gmeiner, Nicolai Marquardt, Michael Bentley, Hugo Romat, Michel Pahud, David Brown, Asta Roseway, Nikolas Martelaro, Kenneth Holstein, Ken Hinckley, and Nathalie Riche. 2025. Intent Tagging: Exploring Micro-Prompting Interactions for Supporting Granular Human-GenAI Co-Creation Workflows. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article 531, 31 pages. doi:10.1145/3706598.3713861 [18] Nitesh Goyal, Minsuk Chang, and Michael Terry. 2024. Designing for Human- Agent Alignment: Understanding what humans want from their agents. In Ex- tended Abstracts of the CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI EA ’24). Association for Computing Machinery, New York, NY, USA, Article 106, 6 pages. doi:10.1145/3613905.3650948 [19]Momchil Hardalov, Arnav Arora, Preslav Nakov, and Isabelle Augenstein. 2022. A Survey on Stance Detection for Mis- and Disinformation Identification. In Findings of the Association for Computational Linguistics: NAACL 2022, Marine Carpuat, Marie-Catherine de Marneffe, and Ivan Vladimir Meza Ruiz (Eds.). Association for Computational Linguistics, Seattle, United States, 1259–1277. doi:10.18653/v1/2022.findings-naacl.94 [20] Kazi Saidul Hasan and Vincent Ng. 2013. Frame Semantics for Stance Classifi- cation. In Proceedings of the Seventeenth Conference on Computational Natural Language Learning, Julia Hockenmaier and Sebastian Riedel (Eds.). Association for Computational Linguistics, Sofia, Bulgaria, 124–132. https://aclanthology. org/W13-3514/ [21] Eric Horvitz. 1999. Mixed-Initiative Interaction. IEEE Intelligent Systems (Septem- ber 1999), 14–24. https://w.microsoft.com/en-us/research/publication/mixed- initiative-interaction/ [22] Yaxin Hu, Anjun Zhu, Catalina L Toma, and Bilge Mutlu. 2025. Designing telep- resence robots to support place attachment. In 2025 20th ACM/IEEE International Conference on Human-Robot Interaction (HRI). IEEE, 252–261. [23]Mads Bruun Ingstrup, Leena Aarikka-Stenroos, and Nillo Adlin. 2021. When institutional logics meet: Alignment and misalignment in collaboration between academia and practitioners. Industrial Marketing Management 92 (2021), 267–276. doi:10.1016/j.indmarman.2020.01.004 [24]A. Karpathy. 2025. Tweet: Vibe coding. Twitter. https://x.com/karpathy/status/ 1886192184808149383 [25]Yubin Kim, Ken Gu, Chanwoo Park, Chunjong Park, Samuel Schmidgall, A. Ali Heydari, Yao Yan, Zhihan Zhang, Yuchen Zhuang, Mark Malhotra, Paul Pu Liang, Hae Won Park, Yuzhe Yang, Xuhai Xu, Yilun Du, Shwetak Patel, Tim Althoff, Daniel McDuff, and Xin Liu. 2025. Towards a Science of Scaling Agent Systems. arXiv:2512.08296 [cs.AI] https://arxiv.org/abs/2512.08296 [26]Robert E. Kraut, Darren Gergle, and Susan R. Fussell. 2002. The use of visual information in shared visual spaces: informing the development of virtual co- presence. In Proceedings of the 2002 ACM Conference on Computer Supported Cooperative Work (New Orleans, Louisiana, USA) (CSCW ’02). Association for Computing Machinery, New York, NY, USA, 31–40. doi:10.1145/587078.587084 [27] Yoonjoo Lee, John Joon Young Chung, Tae Soo Kim, Jean Y Song, and Juho Kim. 2022. Promptiverse: Scalable Generation of Scaffolding Prompts Through Human-AI Hybrid Knowledge Graph Annotation. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (New Orleans, LA, USA) (CHI ’22). Association for Computing Machinery, New York, NY, USA, Article 96, 18 pages. doi:10.1145/3491102.3502087 Alignment–Process–Outcome: Rethinking How AIs and Humans CollaborateCHI EA ’26, April 13–17, 2026, Barcelona, Spain [28] Joseph E. McGrath. 1984. Groups: Interaction and Performance. Prentice-Hall, Englewood Cliffs, NJ. [29]OpenAI. 2024. Introducing OpenAI o1–preview: a reasoning large language model. https://openai.com/index/introducing-openai-o1-preview/. Preview released 12 September 2024; full model released 5 December 2024; Accessed: 2026-01-21. [30] OpenAI. 2025. Agent Builder: A visual canvas for building multi-step agent workflows. https://platform.openai.com/docs/guides/agent-builder. Accessed: 2026-01-21. [31]Michael Polanyi. 1966. The Tacit Dimension. University of Chicago Press, Chicago. [32]Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. 2025. A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications. arXiv:2402.07927 [cs.AI] https: //arxiv.org/abs/2402.07927 [33]Advait Sarkar and Ian Drosos. 2025. Vibe coding: programming through conver- sation with artificial intelligence. arXiv:2506.23253 [cs.HC] https://arxiv.org/ abs/2506.23253 [34] Emanuel Schegloff, Gail Jefferson, and Harvey Sacks. 1977. The Preference for Self-Correction in the Organization of Repair in Conversation. Language 53 (06 1977), 361–382. doi:10.2307/413107 [35]Kjeld Schmidt and Carla Simone. 1996. Coordination mechanisms: Towards a conceptual foundation of CSCW systems design. Comput. Supported Coop. Work 5, 2–3 (Dec. 1996), 155–200. doi:10.1007/BF00133655 [36]Hua Shen, Tiffany Knearem, Reshmi Ghosh, Kenan Alkiek, Kundan Krishna, Yachuan Liu, Ziqiao Ma, Savvas Petridis, Yi-Hao Peng, Li Qiwei, Sushrita Rakshit, Chenglei Si, Yutong Xie, Jeffrey P. Bigham, Frank Bentley, Joyce Chai, Zachary Lipton, Qiaozhu Mei, Rada Mihalcea, Michael Terry, Diyi Yang, Meredith Ringel Morris, Paul Resnick, and David Jurgens. 2025. Position: Towards Bidirectional Human-AI Alignment. arXiv:2406.09264 [cs.HC] https://arxiv.org/abs/2406.09264 [37]Ben Shneiderman and Pattie Maes. 1997. Direct manipulation vs. interface agents. Interactions 4, 6 (Nov. 1997), 42–61. doi:10.1145/267505.267514 [38] Michael Terry, Chinmay Kulkarni, Martin Wattenberg, Lucas Dixon, and Mered- ith Ringel Morris. 2023. Interactive AI Alignment: Specification, Process, and Evaluation Alignment. https://api.semanticscholar.org/CorpusID:264935292 [39] Khanh-Tung Tran, Dung Dao, Minh-Duong Nguyen, Quoc-Viet Pham, Barry O’Sullivan, and Hoang D. Nguyen. 2025. Multi-Agent Collaboration Mechanisms: A Survey of LLMs. arXiv:2501.06322 [cs.AI] https://arxiv.org/abs/2501.06322 [40]Bruce Tuckman. 1965. Developmental sequence in small groups. Psychological Bulletin 63 (06 1965), 384–399. doi:10.1037/h0022100 [41]Raviteja Vemulapalli and Aseem Agarwala. 2019. A Compact Embedding for Facial Expression Similarity. 5676–5685. doi:10.1109/CVPR.2019.00583 [42]Arpita Wadhwa, Aditya Vashistha, and Mohit Jain. 2025. Designing with Culture: How Social Norms Shape Trust and Preference in Health Chatbots. arXiv. https://w.microsoft.com/en-us/research/publication/designing-with- culture-how-social-norms-shape-trust-and-preference-in-health-chatbots/ [43] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, et al.2024. Autogen: Enabling next-gen LLM applications via multi-agent conversations. In First Conference on Language Modeling. [44]Mamehgol Yousefi, Ahmad Shahi, Mos Sharifi, Alvaro J Jorge Romera, Simon Hoermann, and Thammathip Piumsomboon. 2025. Team Dynamics in Human- AI Collaboration: Effects on Confidence, Satisfaction, and Accountability. In Proceedings of the 27th International Conference on Multimodal Interaction (ICMI ’25). Association for Computing Machinery, New York, NY, USA, 395–404. doi:10. 1145/3716553.3750776 [45] J Diego Zamfirescu-Pereira, Richmond Y Wong, Bjoern Hartmann, and Qian Yang. 2023. Why Johnny can’t prompt: how non-AI experts try (and fail) to design LLM prompts. In Proceedings of the 2023 CHI conference on human factors in computing systems. 1–21. [46]Shuning Zhang, Hui Wang, and Xin Yi. 2025. Exploring Collaboration Patterns and Strategies in Human-AI Co-creation through the Lens of Agency: A Scoping Review of the Top-tier HCI Literature. Proc. ACM Hum.-Comput. Interact. 9, 7, Article CSCW413 (Oct. 2025), 43 pages. doi:10.1145/3757594 [47]Xueqiang Zhang, Xiaofei Dong, Yiru Wang, Dan Zhang, and Feng Cao. 2025. A Survey of Multi-AI Agent Collaboration: Theories, Technologies and Appli- cations. In Proceedings of the 2nd Guangdong-Hong Kong-Macao Greater Bay Area International Conference on Digital Economy and Artificial Intelligence (DEAI ’25). Association for Computing Machinery, New York, NY, USA, 1875–1881. doi:10.1145/3745238.3745531 [48] Zeyu Zhang, Quanyu Dai, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2025. A Survey on the Memory Mechanism of Large Language Model-based Agents. ACM Trans. Inf. Syst. 43, 6, Article 155 (Sept. 2025), 47 pages. doi:10.1145/3748302 [49]Zhuohao (Jerry) Zhang, Haichang Li, Chun Meng Yu, Faraz Faruqi, Junan Xie, Gene S-H Kim, Mingming Fan, Angus Forbes, Jacob O. Wobbrock, Anhong Guo, and Liang He. 2025. A11yShape: AI-Assisted 3-D Modeling for Blind and Low- Vision Programmers. In Proceedings of the 27th International ACM SIGACCESS Con- ference on Computers and Accessibility (ASSETS ’25). Association for Computing Machinery, New York, NY, USA, Article 84, 20 pages. doi:10.1145/3663547.3746362 [50]Xinyi Zhou, Zeinadsadat Saghi, Sadra Sabouri, Rahul Pandita, Mollie McGuire, and Souti Chattopadhyay. 2026. Cognitive Biases in LLM-Assisted Software Development. arXiv preprint arXiv:2601.08045 (2026).