Co-agency Collaboration in Mixed Human-AI Teams

Updated 7 July 2026

Co-agency collaboration is a framework where humans and AI systems share initiative, control, and responsibility through negotiated decision-making.
It spans education, scientific research, service design, and embodied systems, emphasizing asymmetry, dynamic role exchange, and contextual authority.
Effective implementation relies on explicit protocols, mutual adaptation, and clear allocation of authority to ensure accountability without over-delegation.

Co-agency collaboration denotes forms of collaboration in which humans and AI systems, or multiple agents, share and negotiate initiative, control, and responsibility while pursuing a common goal. In learning design, it is defined as mixed human–AI teaming that preserves the teacher’s authority over pedagogical intent, constraints, and final decisions (Frøsig et al., 2024). In scientific knowledge production, it is framed as a dynamically reconfigurable partnership with shared epistemic authority (Lin, 6 May 2025), while protocol work generalizes it to heterogeneous autonomous agents whose capabilities are discovered, orchestrated, and governed through explicit communication layers (Liu et al., 18 May 2025). Across these formulations, the central question is not whether AI contributes, but how contribution is bounded, interpreted, and made accountable.

1. Definitions and conceptual scope

Recent work does not treat co-agency as a single doctrine. Instead, it appears as a family of related formulations spanning education, scientific co-creation, service design, accessibility, and embodied shared control. In the educational formulation, co-agency refers to collaboration between teacher and generative AI in which the AI augments the teacher’s capability without displacing the teacher’s power to act, affect matters, make decisions, or take a stance (Frøsig et al., 2024). In Cognitio Emergens, co-agency is a co-evolutionary epistemic partnership in which authority is distributed and continuously adjusted across Directed, Contributory, and Partnership modes (Lin, 6 May 2025). In service co-creation, it is described as the synergistic, shared exercise of agency by humans and LLM-based agents, with shared responsibilities, decision rights, and learning (Zheng et al., 2023). In disability-centered collaboration, the agent is explicitly a bounded partner whose outputs are proposals rather than decisions (Xiao et al., 27 Mar 2026).

A recurrent distinction is that between agency and control. The HCI scoping review defines agency as the capacity to act intentionally and exert power within a context, whereas control denotes the operational means by which intentions are executed and system behavior is influenced (Zhang et al., 8 Jul 2025). This distinction allows co-agency to be analyzed simultaneously at the level of locus of authority, interaction dynamics, and fine-grained mechanisms such as prompting, direct editing, acceptance or rejection, provenance inspection, and adaptive scaffolding.

Domain	Formulation of co-agency	Preserved authority
Learning design	Mixed human–AI teaming in which the teacher and AI collaborate as co-agents	Teacher sign-off and pedagogical intent
Scientific knowledge co-creation	Dynamically reconfigurable partnership with shared epistemic authority	Interpretive integrity
Service co-creation	Synergistic, shared exercise of agency by humans and LLM-based agents	Human-centered boundaries
Disability-centered collaboration	Agent as bounded partner; proposals, not decisions	Human creative authority
Virtual co-embodiment	Joint control over a single avatar	Distributed motor control

This diversity also clarifies a common misunderstanding: co-agency is not identical to equal control. Several frameworks explicitly allow asymmetry, oscillation, and role shifts. Shared agency may involve delegated subtasks, negotiated initiative, or momentary handoffs rather than stable parity (Lin, 6 May 2025).

2. Allocation of authority, initiative, and control

One major line of work analyzes co-agency through explicit dimensions of authority. In the learning-design framework, teacher agency comprises four powers: the power to act, the power to affect matters, the power to make decisions or choices, and the power to take a stance. The hybrid-intelligence arrangement is designed to preserve all four by ensuring teacher initiation, teacher-defined constraints and values, teacher choice among alternatives, and teacher approval before deployment (Frøsig et al., 2024). This turns co-agency into a design problem about checkpoints, editability, provenance, and interruptibility rather than a vague preference for “human oversight.”

Cognitio Emergens makes this allocation formal. If humans are $H=\{h_1,\dots,h_n\}$ and AI agents are $A=\{a_1,\dots,a_m\}$ , authority is represented by a simplex vector $\mathbf{w}(t)\in\Delta^{n+m}$ , with aggregate human and AI shares $W_H(t)$ and $W_A(t)$ . Collaboration mode $C(t)$ can shift among Directed, Contributory, and Partnership states, and these shifts depend on capability signature, governance, and partnership dynamics (Lin, 6 May 2025). The same framework also identifies “epistemic alienation,” namely loss of interpretive ownership over outputs that humans nonetheless endorse. This makes clear that increasing AI participation is not automatically equivalent to stronger co-agency.

A second line of work emphasizes mutual adaptation rather than static allocation. The co-learning framework describes human and AI as “two dynamic, learning entities” whose collaboration depends on mutual understanding, mutual benefits, and mutual growth (Huang et al., 2019). The scoping review on co-learning and co-adaptation distinguishes “co-adaptation” as reciprocal adjustment in behavior or control policies during interaction and “co-learning” as the longer-horizon joint change in knowledge, strategies, and shared mental models (Kumar et al., 30 May 2025). This suggests that co-agency can be episodic and task-bounded, or developmental and longitudinal, depending on whether the collaboration chiefly redistributes control or also transforms the partners’ capabilities.

The HCI literature further decomposes co-agency into locus, dynamics, and granularity. Locus concerns whether authority is human-centric, AI-centric, or hybrid; dynamics concerns whether roles are static or dynamically renegotiated; granularity concerns whether agency is exercised at goal level or parameter level (Zhang et al., 8 Jul 2025). In practice, durable co-agency tends to require simultaneous support at both macro and micro levels: high-level goal framing, plus low-level opportunities to revise, reject, or redirect AI behavior.

3. Educational and creative instantiations

In educational technology, co-agency is often articulated as a correction to learner-centric systems that under-specify teacher authority. The hybrid-intelligence framework for learning design assigns pedagogical intent, contextual knowledge, ethical judgment, and final decision rights to the teacher, while generative AI contributes rapid idea generation, drafting, tailoring to level, pattern detection, option synthesis, and rationale presentation when explainability is available (Frøsig et al., 2024). The collaboration workflow is modular: the teacher sets goals and constraints; the AI proposes multiple options annotated with rationales and uncertainty; teacher and AI co-create artifacts such as lesson outlines, question sets, and rubrics; the AI pre-screens for alignment, bias, and readability; and final approval remains with the teacher before classroom use. The architecture is explicitly designed to prevent one-shot opaque automation such as “generate a lesson plan” from narrowing the teacher’s power to act.

Child–AI co-creation in Tinker Tales operationalizes a similar principle under stronger pedagogical scaffolding. Children remain the primary decision-makers: they select characters and stage-specific place, item, and emotion tokens, then co-develop the narrative through voice dialogue. The AI asks scaffolded narrative and social-awareness questions, integrates the child’s additions into updated story segments, and visibly carries forward those inputs. In the exploratory study, Primitive narratives prompts elicited event additions in 90% of responses, Chain narratives prompts elicited contributions in 100% of responses, and Social awareness prompts elicited input in 62% of responses. AI uptake of child contributions was Full in 90% of cases and Partial in 10%, with No uptake not observed (Choi et al., 4 Feb 2026). The design claim is not that the AI leads, but that scaffolding can increase coherence without diminishing child agency.

Practice-based design research reaches a related conclusion from the opposite direction. In textile-making experiments with Bard, designers initially elevated the model to a superior guide and experienced loss of creative agency through repetition, misinterpretation, and failure to honor material constraints. Agency was regained when designers reframed Bard as assistant or suggester, reduced prompt volume, imposed clearer constraints, and explicitly rejected incorrect outputs (Lin et al., 12 Mar 2026). The reported stages—Questions & Feedback Overload, Repetition of Reminders, Communication Development, Stable Communication—show that co-agency in creative work is often less a steady state than a recurrent renegotiation of role assignment and dependence.

A broader creativity taxonomy places these cases on a continuum. “Digital Pen” tools merely digitize practice; “AI Task Specialist” systems automate bounded subtasks; “AI Assistant” systems support interactive ideation; and “AI Co-Creator” systems contribute non-intuitive proposals within a tight propose–evaluate–refine loop (Haase et al., 2024). This suggests that co-agency is strongest not when AI autonomy is maximal, but when initiative is shareable and evaluative authority remains legible.

4. Multi-agent coordination, auditing, and protocolized collaboration

Co-agency is not limited to human–AI dyads. In fairness auditing, multiple agents can collaborate to estimate demographic parity across different protected attributes under a fixed query budget. The formal result is that collaboration is generally beneficial relative to independent audits, except for a-priori collaboration with equal stratification across the full $2^m$ intersectional strata, which can become disadvantageous as the number of agents grows (Vos et al., 2024). Under a-posteriori collaboration with uniform sampling, the variance satisfies

$Var(\hat{D})^{uniform}_{a\text{-}posteriori}=\frac{1}{m}Var(\hat{D})^{uniform}_{no\ collab},$

and the experiments on Folktables, German Credit, and ProPublica COMPAS showed convergence of uniform, stratified, and Neyman performance as $m$ rises, with the Neyman–stratified gap reported as very small $(<0.001)$ (Vos et al., 2024). The important controversy here is that “more coordination” is not unconditionally better: extensive a-priori coordination on exponentially many strata can degrade accuracy.

Protocol work makes the same issue infrastructural. ACPs defines co-agency in the Internet of Agents through a division among Initiator, Orchestrator, and Responders, supported by Agent Registration Protocol, Agent Discovery Protocol, Agent Interaction Protocol, and Agent Tooling Protocol (Liu et al., 18 May 2025). The model assumes heterogeneous agents, autonomy in perception/decision/action, and cross-platform identity and security. Orchestration handles task decomposition, capability matching, routing, and aggregation; choreography remains possible through agent-to-agent negotiation. In this framing, co-agency depends on reliable capability advertisement, trusted identity, tool access scoping, session security, and accountable workflow construction.

AWCP addresses a narrower but consequential gap: message-passing does not grant a remote agent direct access to a peer’s live execution environment. It therefore introduces temporary workspace delegation, in which a Delegator projects a scoped directory tree to an Executor through a control plane and pluggable filesystem-level transports (Nie et al., 24 Feb 2026). The lifecycle—INVITE, ACCEPT, START, DONE, cleanup—formalizes deep-engagement collaboration on shared files rather than mere exchange of JSON payloads. This is a technically distinct form of co-agency: the collaborating agents do not only coordinate intentions; they operate within the same workspace.

Ad hoc teamwork in Avalon shows the converse case, where collaboration lacks stable protocols. LLM agents must infer teammate types and intentions from partial observations and noisy natural-language interaction. Communication did not significantly improve GPT-4 self-play performance: with communication, Game Win was 0.500, Quest Win 0.513, and Team Selection Accuracy 0.640; without communication, the corresponding values were 0.466, 0.480, and 0.693 (Shi et al., 2023). CodeAct improved team selection by coupling enhanced memory with code-driven reasoning, yielding Team Acc 0.830 versus 0.707 for CoT and 0.634 for ReAct (Shi et al., 2023). A plausible implication is that co-agency in language-based teams depends less on the sheer presence of dialogue than on factual memory, verifiable reasoning, and resistance to hallucination.

5. Embodied, situated, and ability-diverse collaboration

Embodied settings make co-agency directly perceptual. In virtual co-embodiment, two users share control of a single avatar through weighted fusion,

$A=\{a_1,\dots,a_m\}$ 0

with discrete control weights such as W0, W25, W50, W75, and W100 (Fribourg et al., 2019). Experimental results show that participants track their actual control well, but significantly overestimate their sense of agency when they can anticipate avatar motion. In the Target task, even W0 produced higher Feeling of Control than in Free and Trajectory, and motion offsets were strongly negatively correlated with FoC across tasks, with $A=\{a_1,\dots,a_m\}$ 1, $A=\{a_1,\dots,a_m\}$ 2, and $A=\{a_1,\dots,a_m\}$ 3 respectively (Fribourg et al., 2019). Co-agency here is partly prospective: shared intention and predictable movement can inflate perceived authorship.

ShareYourReality extends this by adding position-aware vibrotactile feedback during avatar co-embodiment. The study found lower SoA in the free-choice task with haptics than without, higher SoA in the shared targeted task, significantly higher co-presence in free-choice, and stronger hand-motion synchrony in the targeted task (Venkatraj et al., 2024). The result is a caution against assuming that more feedback automatically strengthens collaboration. In free-choice settings, overlap-based haptics can introduce attribution ambiguity rather than coordination support.

Situated workplace research shifts the focus from shared control to socio-material presence. In co-located blue-collar work, embodied AI agents are proposed as visible and audible team members that support shared situational awareness, hands-free interaction, and inclusive communication across experience levels (Vaananen et al., 12 Feb 2026). The smart-factory scenario involves a physically present agent that analyzes sensor data, suggests causes, and adapts its troubleshooting model based on the team’s chosen solution. The maintenance scenario uses a transcription agent that overhears expert–novice dialogue, structures documentation in the background, and builds a shared knowledge base for later learning. The key design claim is that embodiment should be treated as a socio-material design strategy rather than an aesthetic add-on.

The disability-centered framework generalizes this situated view into three nested layers: Channelling, Coordinating, and Co-Creating (Xiao et al., 27 Mar 2026). Channelling establishes shared informational ground through modality-adapted representations; Coordinating mediates workflow, handoffs, and awareness; Co-Creating positions the agent as a bounded partner whose outputs remain transparent, modifiable, and clearly distinguished from human-authored work. This explicitly rejects the individual-assistance model in favor of triadic, ability-diverse collaboration in which access is produced through interdependence rather than assumed as an individual property.

6. Evaluation, governance, and recurrent tensions

Evaluation of co-agency is notably multi-dimensional. In learning design, proposed measures include perceived control, decision influence, outcome alignment, trust calibration, workload and time, quality and fairness, reliance versus learning, and algorithm aversion resilience (Frøsig et al., 2024). Cognitio Emergens similarly rejects scalar performance alone and proposes a composite metric

$A=\{a_1,\dots,a_m\}$ 4

where weights vary with configuration and components may include interpretability, robustness, provenance integrity, accountability, adaptability, knowledge transfer, and transformative potential (Lin, 6 May 2025). The methodological consequence is that co-agency cannot be reduced to accuracy, throughput, or user satisfaction in isolation.

Governance mechanisms recur across domains. In education they include teacher initiation, constraint authoring, sign-off, opt-in or opt-out data use, explainability, provenance, and accountability logs (Frøsig et al., 2024). In service co-creation they include knowledge boundaries, source inspection, escalation pathways, policy-mode verbatim handling, disclaimers, and privacy management distinct from human operators (Zheng et al., 2023). In accessibility-centered collaboration they include explicit consent for sensing, visibility of collaboration states, pause and override controls, provenance tags, and the norm that AI contributions are proposals rather than decisions (Xiao et al., 27 Mar 2026). Across protocol work, trusted identity, authorization, and auditable usage occupy the same role at system scale (Liu et al., 18 May 2025).

Several tensions recur with remarkable consistency. One is opacity: black-box outputs reduce the human capacity to affect matters, take a stance, or calibrate trust (Frøsig et al., 2024). Another is over-delegation: epistemic alienation in scientific collaboration, deskilling in design practice, and over-reliance in educational contexts each describe cases where endorsement outpaces interpretation (Lin, 6 May 2025). A third is that visible process is not sufficient by itself. In concurrent co-creative interaction, process visibility helped designers reason about agent actions, but conflicts emerged when the agent could not distinguish feedback from independent work; this motivated context-aware attribution and plan updating in CLEO (Son et al., 2 Mar 2026). The empirical distribution in Study 2—Hands-off in 70.1% of turns, Concurrent work in 31.8%, Direction in 28.5%, and Termination in 8.9%—shows that co-agency includes dynamic switching rather than a single “collaboration mode” (Son et al., 2 Mar 2026).

Terminology itself remains unsettled. The scoping review of co-learning and co-adaptation notes that the prefix “co” is used interchangeably for “collaborative” and “mutual,” and that “co-learning” and “co-adaptation” are sometimes conflated despite emphasizing different timescales and mechanisms (Kumar et al., 30 May 2025). This suggests that co-agency collaboration is best treated not as a monolithic theory but as a structured design problem: how to distribute initiative, make control contestable, preserve interpretive and moral responsibility, and support adaptation without silent capture of human judgment.

In sum, co-agency collaboration is a technical and organizational arrangement for sharing agency under conditions of asymmetry, uncertainty, and evolving competence. Whether the setting is teacher–genAI learning design, multi-agent fairness auditing, scientific knowledge production, shared-avatar embodiment, blue-collar teamwork, or disability-centered collaboration, the decisive variables are explicit decision rights, transparent contribution traces, bounded autonomy, and mechanisms for correction, renegotiation, and final human accountability (Frøsig et al., 2024).