The sustained coordination of multiple actors toward shared outcomes, transforming independent efforts into joint accomplishments without confusion, duplication, or loss of individual accountability.

See Collaboration foundation for conceptual grounding.

Variants

Temporal

• Real-time collaboration: simultaneous participation with immediate feedback
• Asynchronous collaboration: time-shifted contributions with integrated workflows
• Near-synchronous collaboration: hybrid approach with delayed but frequent updates

Participant types

• Human ↔ Human
• Human ↔ Agent

Interaction stages

These stages describe the shared workflow actors move through when creating together, from initial perception to final evaluation.

1. Perceive: actors gather and interpret information to establish a shared understanding of the task context and goals.
2. Think: This stage encompasses the internal cognitive or computational processes where both human and AI formulate ideas, strategies, and potential solutions.
3. Express: Articulating and externalizing internally generated ideas or results. The clarity, relevance, and interpretability of expression are crucial for collaborator comprehension and subsequent action, influencing the perceived usefulness and agency of each participant.
4. Collaborate: The core interactive partnership, characterized by iterative exchanges aimed at refining ideas, resolving discrepancies. Involves turn-taking, negotiation, critique, modification, and coordinated responses. Effective collaboration requires mechanisms that foster mutual understanding, transparency regarding actions and reasoning, and well-designed communication protocols.
5. Build: Translating collaboratively refined concepts into tangible artifacts or implemented solutions. Contributions from earlier stages serve as foundational blocks, templates, or reference materials. Actors may work in parallel or sequentially.
6. Test: Evaluating the co-created output against initial goals or emergent criteria. Actors assess quality, functionality, and alignment with shared vision. Evaluation often involves providing feedback through various modalities (explicit adjustments, natural language corrections, ratings), which can loop back to earlier stages for iteration.

Human ↔ Human collaboration

Real-time collaboration

Enables users to work together simultaneously, with immediate feedback and shared presence.

TODO: Live cursors, presence indicators, messaging, simultaneous editing

Asynchronous collaboration

Allows users to contribute at their own pace, without needing to be online at the same time.

TODO: Commenting, annotations, notifications, activity log, version management

…

Human ↔ Agent

Human-AI collaboration unfolds through the interaction stages with shifting agency patterns, distribution, and control mechanisms.

Perceive

Actors gather and interpret information to establish shared understanding of the task context and goals. Actors formulate intent and provide input.

• Agency: typically reactive—agent responds to user input without independent initiative
• Distribution: human-centric locus with user providing initial direction and framing
• Control mechanisms: guided input structures user prompts and parameters; context awareness enables agent to leverage history and environment

Think

Internal cognitive or computational processes where actors formulate ideas, strategies, and potential solutions. This involves processing perceived input, generating novel representations, planning approaches, and evaluating possibilities. The alignment (or misalignment) between collaborators’ thinking critically impacts the trajectory of collaboration.

• Agency: semi-active to proactive—agent analyzes inputs and may generate insights or suggestions independently
• Distribution: shifting toward shared locus as agent contributes substantive thinking; dynamic allocation based on problem complexity
• Control mechanisms: transparency makes agent reasoning visible; intervention allows real-time adjustment; context management maintains task continuity

Express

Articulating and externalizing generated ideas or results. Actors manifest thoughts through various outputs: text, multimodal content, actions, or decisions. The clarity, relevance, and interpretability of expression are crucial for collaborator comprehension and subsequent action, influencing the perceived usefulness and agency of each participant.

• Agency: reactive to semi-active—agent generates outputs in response to prompts or contextual triggers
• Distribution: varies by output type; may be human-centric (user directs generation) or shared (agent proposes options)
• Control mechanisms: transparency distinguishes AI-generated content; adaptive scaffolding adjusts output complexity; modification enables direct editing

Collaborate

The core interactive partnership, characterized by iterative exchanges aimed at refining ideas and resolving discrepancies. Involves turn-taking, negotiation, critique, modification, and coordinated responses. Agency becomes highly distributed and dynamically negotiated during this stage. Effective collaboration requires fostering mutual understanding and transparency regarding actions and reasoning.

• Agency: co-operative—agent acts as active partner, both actors mutually influence the creative output
• Distribution: shared locus with dynamic allocation as control shifts based on task phase; both high-level (strategic) and fine-grained (tactical) granularity
• Control mechanisms: full range applies—iterative feedback loops enable refinement; intervention allows course correction; transparency maintains shared understanding; confidence visualization surfaces reliability

Build

Translating collaboratively refined concepts into tangible artifacts or implemented solutions. Contributions from earlier stages serve as foundational blocks, templates, or reference materials. Actors may work in parallel or sequentially, with agency pivoting towards execution, integration, and fine-grained control during materialization of the co-created vision.

• Agency: proactive to co-operative—agent may execute tasks autonomously or work alongside user
• Distribution: may shift between human-centric, shared, and system-centric based on sub-task; fine-grained control during implementation
• Control mechanisms: intervention enables real-time adjustments; modification allows direct editing of outputs; adaptive scaffolding adjusts assistance levels; chain-of-thought reveals execution reasoning

Test

Evaluating the co-created output against initial goals or emergent criteria. Actors assess quality, functionality, and alignment with shared vision. Evaluation often involves providing feedback through various modalities (explicit adjustments, natural language, ratings, embodied reactions), which can loop back to earlier stages for iteration.

• Agency: semi-active—agent provides evaluation support and identifies issues, but doesn’t make final judgments
• Distribution: human-centric locus for ultimate acceptance/rejection decisions; agent contributes analysis
• Control mechanisms: confidence visualization indicates agent’s certainty about issues; explanatory feedback clarifies evaluation reasoning; iterative feedback loops enable refinement cycles; intervention allows override of agent assessments

Human elaboration of agent contributions

Level 1-2: Surface acceptance

Human accepts or minimally modifies AI output without substantial original contribution.

• Lowest cognitive engagement and perceived ownership
• Risk of algorithmic loafing—reduced human effort when AI takes creative lead
• Appropriate for routine tasks but problematic for creative or learning contexts

Level 3: Substantive elaboration

Human adds independent features, mechanisms, constraints, specifications, or context not provided by AI.

• Moderate-high cognitive engagement
• Concrete detailing that extends AI contributions
• Maintains human creative agency

Level 4: Integrative advancement

Human creates incrementally new ideas by combining, syncing, or restructuring components from AI input to extend or refine the core concept.

• Requires active synthesis between human and AI contributions
• High cognitive engagement and co-creative ownership

Level 5: Human-led reframing

Human pivots core concept to serve different primary purpose, audience, or context—replacing the core frame whilst leveraging AI input as inspiration.

• Maximum cognitive engagement and perceived ownership
• Radical innovation through conceptual pivoting
• Demonstrates human creative lead with AI as scaffolding

Factors influencing elaboration

• Agency distribution: Human-centric locus enables deeper elaboration; system-centric locus often correlates with surface acceptance
• Interface affordances: What modifications the interface permits
• Cognitive capacity: Available mental resources and task complexity
• Task nature: Creative vs routine work; exploratory vs optimisation goals

Cognitive effort paradox

Higher cognitive workload during collaboration correlates with increased perceived ownership and deeper elaboration (Maier et al. 2025). Whilst reducing mental effort seems beneficial, appropriate cognitive load maintains skill development, creative engagement, and protection against algorithmic loafing.

Elaboration across collaboration stages

• Perceive/Think: Higher elaboration maintains human agency in problem framing
• Express: Surface acceptance risks fixation on AI-provided options
• Collaborate: Substantive elaboration enables shared locus and mutual influence
• Build: Elaboration depth from earlier stages carries forward into execution quality
• Test: Higher elaboration enables more critical evaluation and iteration

See Conversation for how dialogue forms shape elaboration in conversational interfaces, and Agent for interaction patterns that maintain engagement.

Agent ↔ Agent

…

Resources & references

• Maier, Schneider, Feuerriegel (2025) Partnering with Generative AI: Experimental evaluation of human-led and model-led interaction in human-AI co-creation
• Zhang, Wang, Yi (2025) Exploring Collaboration Patterns and Strategies in Human-AI Co-creation through the Lens of Agency
• Nardi et al (1998) Collaborative, programmable intelligent agents
• Matt Webb, Designing multiplayer apps with patterns from architecture
• Apple HIG