Recoleta
GitHub
Source note

Building Persona-Based Agents On Demand: Tailoring Multi-Agent Workflows to User Needs

Authors
Giuseppe Arbore, Andrea Sillano, Luigi De Russis
Source
arXiv
Published
Collection
Software Intelligence
Multi Agent SystemsPersona AgentsAgent OrchestrationPersonalizationHuman AI Interaction
Open arXivMarkdown source

Summary

The paper proposes a runtime pipeline that creates persona-based agents for each user query, using the user profile, task needs, and session context. It targets personalized multi-agent workflows, but the excerpt reports no implementation, benchmark, or user study.

Problem

  • Multi-agent systems often use preset roles, communication paths, and execution order, which makes them hard to adapt to user expertise, task context, and preferred interaction style.
  • This matters because users may need different explanations, task splits, and agent behaviors for the same high-level request.
  • Fixed agent setups require prompt or orchestration changes when the user or task changes.

Approach

  • A central orchestrator receives an open-ended query and runs ProfileEncode to infer user traits, intent, domain familiarity, and communication preferences.
  • TaskDecompose splits the query into subtasks and records dependencies so some tasks can run in parallel while others wait for required outputs.
  • PersonaCraft creates a persona for each subtask, with role, domain skills, communication style, and capabilities drawn from the query and user profile.
  • AgentFactory instantiates an LLM-backed agent from each persona, then the orchestrator assigns tasks, routes intermediate outputs, and aggregates the final answer.
  • The system keeps generated personas and agents within a session and can add new ones on later queries.

Results

  • The excerpt gives 0 quantitative evaluation results: no accuracy, task-completion, latency, cost, or user-satisfaction numbers are reported.
  • The claimed system flow has 4 stages: Query Analysis, Agent Generation and Instantiation, Agent Assigning and Execution, and Answers Aggregation and Displaying.
  • The algorithm lists 18 steps for on-demand persona-based agent generation, starting with a user query and ending with an aggregated response.
  • The strongest concrete claim is architectural: agent roles, styles, and coordination behavior are generated at runtime from the user profile, task plan, and session context.
  • The paper claims the session model improves coherence across multiple queries by retaining prior agents and personas, but it provides no measured comparison against fixed-role multi-agent systems.

Link