Channelling, Coordinating, Collaborating: A Three-Layer Framework for Disability-Centered Human-Agent Collaboration

People with disabilities are increasingly using LLM-based tools across their everyday workflows, from information access [1] and day-to-day decision support [23] to end-to-end content creation [26]. However, AI accessibility tools and much of the surrounding HCI research remain focused on individual assistance: compensating for a specific functional limitation through a one-person, one-tool interaction model. Disability scholarship has long argued that this framing is incomplete. Bennett et al. [6] argue that interdependence, not independence alone, better captures how people with disabilities navigate daily life, and should sit at the centre of assistive technology research. Xiao et al. map recurring patterns in ability-diverse collaboration, showing that collaboration is itself the mechanism through which access gets produced [25]. This collaborative reality is evident across settings: blind runners coordinate with sighted guides through continuous, fine-grained communication that weaves together voice cues and bodily movement [5], and creators with sensory impairments split the work of video production across multiple stages with trusted partners, from filming through editing and publishing [26]. For many people with disabilities, collaboration is the ordinary infrastructure behind complex professional and creative work. Human-agent collaboration research offers relevant foundations for understanding this collaborative reality, with frameworks for trust, coordination, and shared agency [2, 15, 18, 28] building on grounding theory [10], workspace awareness [14, 16], and knowledge boundary management [9]. Yet these frameworks have not engaged with ability-diverse settings, and carry assumptions that need rethinking when ability differences are present. In this position paper, we propose a three-layer framework that bridges accessibility and human-agent collaboration. Layer 1, Channelling, concerns modality-adapted, equivalent access to task-relevant information. Layer 2, Coordinating, concerns how AI can mediate workflows, communication, and handoffs among collaborators with different abilities. Layer 3, Co-Creating, moves beyond support to consider AI as a bounded contributor that advances shared goals alongside human partners. Grounded in the Ability-Diverse Collaboration framework [25], grounding theory [10], and Carlile’s 3T framework [9], together these layers extend both fields by centring interdependence in people with disabilities’ everyday practices and by treating triadic collaboration as an explicit design target.

2 Background and Positioning Ability-diverse collaboration, where collaborators bring fundamentally different abilities to shared work, differs from typical collaboration in two key ways [25]. First, collaborators may not enjoy equal access to information, nor possess congruent knowledge concerning the content and process of collaboration. Second, this inherent asymmetry in abilities and information access can engender distinct roles and potentially divergent goals. These two features challenge foundational assumptions in the theories most commonly used to explain collaboration. Clark and Brennan’s [10] theory of grounding describes how collaborators establish mutual understanding through a collective process shaped by the communication medium’s constraints: copresence, visibility, audibility, cotemporality, and others. A central principle is least collaborative effort: participants minimise the total work needed to reach mutual understanding. In ability-diverse collaboration, however, these constraints are not properties of the medium alone but of the person-medium coupling. A video call affords visibility, but not for a blind collaborator. A voice channel affords audibility, but not for a Deaf collaborator. The first feature of ability-diverse collaboration, unequal information access, means that grounding costs become fundamentally asymmetric. The leasteffort path for the group may concentrate information gatekeeping in one party, efficient in aggregate but constraining for the person with a disability. Carlile’s [9] 3T framework describes three progressively complex knowledge boundaries: syntactic (a shared lexicon suffices), semantic (interpretive differences require translation), and pragmatic (conflicting interests mean knowledge is “at stake” and must be transformed). The second feature of ability-diverse collaboration, that asymmetry creates distinct roles and divergent goals, maps directly onto these pragmatic boundaries. The knowledge at stake is not only domain-specific expertise but ability-specific: the embodied, situated understanding of how to perceive, move, and work. When a collaboration must be renegotiated because abilities or roles change, this knowledge carries real costs, and mismatches between the boundary type and the process used produce costly failures. Human-agent collaboration research has been exploring agents as collaborative partners, building on grounding theory and workspace awareness [14, 16], and developing frameworks for trust, coordination, and shared agency [2, 18, 28]. This creates a natural opportunity: AI could address both features of ability-diverse collaboration, reducing information asymmetry and mediating the distinct roles that ability differences create. However, current frameworks presume shared perceptual access to the same evidence, a common modality for presenting awareness information, and a dyadic structure. In ability-diverse settings, AI typically enters an existing partnership, creating a triad with fundamentally different coordination demands. Addressing these challenges requires a framework that accounts for both: how AI can reconfigure the grounding conditions so that information access is no longer dependent on a single collaborator, and how AI can enter the negotiation of roles and goals without overriding the agency of the people involved.