Towards AI as Colleagues: Multi-Agent System Improves Structured Ideation Processes

Paper Title

Towards AI as Colleagues: Multi-Agent System Improves Structured Ideation Processes

Publication Info

• Topic area: Human–AI collaboration in creative ideation using multi-agent systems.
• Keywords: AI collaboration, multi-agent systems, ideation processes, human–AI teaming, divergent thinking, convergent thinking, creativity support, role differentiation, facilitation, GPT-4o.

Background and Problem

• Problem / challenge: Current AI systems primarily function as tools or assistants, limiting their ability to act as collaborative partners in creative ideation. Single-agent systems often lack diversity in perspectives and structured interaction dynamics.
• Significance: Enhancing AI systems to act as colleagues could improve creativity, expand idea exploration, and foster more collaborative human–AI interactions.
• Motivation and related work: Prior research has explored multi-agent frameworks and persona-based role differentiation, but these approaches often remain fragmented and lack integration of real-world brainstorming dynamics. This paper builds on these foundations to address gaps in collaborative ideation systems.

Solution

• Proposed approach: MultiColleagues, a multi-agent conversational system that integrates diverse AI personas with structured facilitation and adaptive thinking modes to support human–AI co-ideation.
• Novelty:
  1. Dynamic turn-taking and role differentiation among AI personas to emulate team dynamics.
  2. Integration of divergent and convergent thinking modes inspired by the double diamond design methodology.
  3. Human-centered facilitation to maintain oversight and guide transitions between ideation phases.
  4. User-friendly interfaces for managing multi-agent collaboration and cognitive load.
• Procedure and key techniques:
  • Adaptive thinking transitions: Explore mode for broad idea generation and Focus mode for refinement and synthesis.
  • Persona orchestration: AI colleagues with distinct roles, communication styles, and expertise contribute in structured turns.
  • Facilitation mechanisms: Real-time monitoring and intervention to guide discussion direction and prevent cognitive overload.
  • Interaction design: Visual cues, role-based message styling, and controls for switching thinking modes.

Results

• Concrete findings:
  • MultiColleagues elicited nearly twice as many user contributions compared to the single-agent baseline (M = 8.35 vs. M = 4.10 utterances, p = .001).
  • Higher ratings for perceived outcome quality and novelty (M = 5.95 vs. M = 4.97, p < .01).
  • Enhanced engagement and flow (M = 5.70 vs. M = 4.45, p = .014).
• Advantage over baselines:
  • MultiColleagues fostered stronger team-like feelings and complementary strengths across roles.
  • Broader and deeper exploration with higher topic branching (M = 6.45 vs. M = 3.50, p < .001) and concept production rates (M = 2.05 vs. M = 1.26 concepts/min, p < .05).
• Experiments / evaluation:
  • Within-subjects study with 20 participants comparing MultiColleagues to ChatGPT (single-agent baseline).
  • Metrics included user ratings, linguistic cohesion, pragmatic style, and originality scoring.
  • Mixed-methods approach integrating surveys, interviews, and conversation log analyses.
• Limitations and future work:
  • Limited participant diversity (students and early-career professionals).
  • Short interaction sessions (10 minutes); longer-term studies needed.
  • Use of homogeneous language models; future work should explore heterogeneous setups and bias auditing.

Summary

MultiColleagues demonstrates how multi-agent systems can shift AI from tools to peer-like collaborators in creative ideation. By combining role differentiation, adaptive thinking modes, and facilitation, the system fosters broader exploration, stronger engagement, and higher perceived outcome quality compared to single-agent workflows. Results highlight the potential of multi-agent frameworks to emulate team dynamics and support structured creativity. Future work should address limitations in participant diversity, interaction duration, and model heterogeneity to further refine AI colleagueship.