Conversation as a Computational Instrument
A Structured Multi-Agent Framework for Discovering Individual and Group Decision Needs
Abstract
Most AI decision-support systems rely on static profiles, predefined preferences, or single-pass reasoning to generate recommendations. Multi-agent systems and generative agent simulations exist and successfully demonstrate that agents can exhibit realistic social behavior, memory, and emergent interaction over time. However, these systems primarily treat conversation as an outcome to be observed for realism, coherence, or narrative quality.
This work proposes a different paradigm in which structured conversation itself becomes the primary computational instrument for discovering what both a group and its individual members need to understand and decide next.
By transforming interaction into structured insights and targeted follow-up questions, the system creates an iterative loop of understanding and decision formation that cannot be achieved through flat AI analysis or open-ended agent simulation alone.
1. Background and Related Work
Existing approaches fall into three main categories:
Flat AI decision-support systems
These systems analyze structured inputs (profiles, preferences, constraints) and output decisions or recommendations. They assume that relevant information is already explicit and stable.
Multi-agent debate and deliberation systems
These systems simulate argumentation or reasoning among agents to converge on correct or optimal outcomes. Agents represent positions or roles rather than human identities, and the primary objective is solution quality rather than interaction dynamics.
Generative agent simulations (e.g., Stanford Generative Agents)
These systems explore emergent social behavior and realism in simulated environments. Their goal is narrative coherence and behavioral plausibility, not structured discovery of decision-relevant knowledge.
None of these frameworks assigns conversation a functional role as a diagnostic or computational layer for discovering what is missing, unclear, or unresolved in human understanding and decision-making.
2. Core Idea
We propose a system in which:
- Agents are grounded in strict human profiles and behavioral constraints.
- A conversation is generated for a given task (e.g., planning, leadership selection, conflict resolution).
- The system produces a structured report summarizing interaction dynamics such as alignment, disagreement, uncertainty, unresolved issues, and changes in participant feedback over time.
- A user may propose a follow-up task (e.g., “let’s vote for a leader” or “let’s plan a trip together”).
- The system generates targeted questions for each participant based on what the conversation revealed as missing or ambiguous.
- Participant answers update structured profiles.
- A new conversation is generated using the enriched profiles and new task context.
This creates the loop:
Conversation → Insight → Structured Questions → Profile Update → New Conversation
3. Novelty Claim
The novelty of the proposed framework lies not in multi-agent conversation itself, but in the functional role assigned to conversation.
Conversation is not treated as a user interface or a narrative artifact. Instead, it is treated as a computational instrument for discovering decision-relevant knowledge that cannot be derived from static profiles or flat analysis.
Key distinguishing features
Conversation as a diagnostic layer
Interaction reveals priorities, tensions, misunderstandings, and value conflicts that are not observable from structured inputs alone.
Automatic generation of new structured knowledge
The system transforms interaction outcomes into targeted follow-up questions, producing new structured inputs for subsequent reasoning.
Dual-level discovery
The system identifies:
- what the group needs to decide next
- what each individual member still needs to understand or clarify
Iterative epistemic process
Each conversation modifies the informational state of the system, enabling progressively deeper and more accurate subsequent interactions.
4. Why Flat AI Is Insufficient
Static AI systems assume that preferences are known, goals are explicit, and conflicts are directly representable in data.
Human decision-making is fundamentally interactional:
- people discover what matters by responding to others,
- misunderstandings emerge only through dialogue,
- alignment and tension become visible only through interaction.
Therefore, conversation generates information that does not exist prior to the interaction and cannot be inferred from static inputs alone.
5. Contribution
This work introduces a new interaction paradigm for AI-supported understanding and decision systems:
A conversation-driven discovery framework in which structured interaction generates the next questions to be asked, and user-selected follow-up tasks are translated into targeted questions and a new conversation.
The framework shifts AI from:
analyzing people → enabling people (or agents) to reveal themselves through interaction.
6. Conclusion
We argue that conversation can function as a computational instrument for discovering decision-relevant knowledge. By structuring conversation outputs into reports and question generation, the proposed framework enables iterative refinement of both individual understanding and group decision-making.
This establishes a new paradigm in which interaction itself becomes the engine of knowledge generation, rather than merely a channel for communication.