Skip to content

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

🕒 Published (v1): 2026-04-30 14:01 UTC · Source: Arxiv · link

Why this paper was selected

Persona-based agent workflow tailoring on demand; adaptive multi-agent customization

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This position paper proposes a pipeline for generating LLM-backed agent personas on demand at runtime, replacing fixed hard-coded agent architectures in multi-agent systems. An orchestrator dynamically synthesizes roles, communication styles, and tool-use strategies tailored to each user query and user profile. The work is a design proposal without empirical evaluation.

Problem

Current LLM-based multi-agent systems encode agent roles, coordination patterns, and interaction protocols as design-time constants, making it difficult to adapt to individual users, varying expertise levels, or diverse task contexts without manually revising prompts or orchestration logic.

Method

A session-scoped orchestrator executes four sequential steps upon each user query:

  1. ProfileEncode — extracts implicit/explicit signals (domain expertise, communication style, task familiarity, intent) from the raw query to build a user profile \(u\).
  2. TaskDecompose — decomposes the query into a dependency-aware task graph \(\{t_1, \ldots, t_n\}\), annotating which tasks are parallelizable vs. sequential.
  3. PersonaCraft + AgentFactory — for each task \(t_i\), synthesizes a persona specification \(p_i\) (role, domain competencies, communication style, capabilities) conditioned on \(t_i\), the user profile \(u\), and already-existing personas (to avoid duplication), then instantiates an LLM-backed agent \(a_i \leftarrow \text{AgentFactory}(p_i)\).
  4. Answers Aggregation — collects partial results \(\{r_1, \ldots, r_n\}\) from agents executed in dependency order, merges them, resolves inconsistencies, and aligns output style with user preferences.

Within a session, generated agents and personas are retained and extended across follow-up queries rather than discarded.

Key Contributions

  • Framing of on-demand persona generation as a first-class architectural primitive alongside memory, tool use, and orchestration in agentic platforms.
  • A concrete four-step pipeline (ProfileEncode → TaskDecompose → PersonaCraft/AgentFactory → Aggregate) that moves agent roles from design-time constants to runtime variables.
  • A session-scoped model that preserves generated agent pools across multi-turn interactions while allowing incremental persona creation.
  • Positioning persona-conditioned generation as a mechanism for end-user empowerment across heterogeneous expertise levels and task contexts.

Results

No empirical results are reported. This is a position/workshop paper presenting a conceptual pipeline without quantitative evaluation.

Limitations

  • No implementation or empirical evaluation; all claims are theoretical.
  • No analysis of failure modes when ProfileEncode produces an inaccurate user model, which could propagate errors through the entire pipeline.
  • Session-scoped memory is deliberately discarded between sessions, preventing long-term user adaptation.
  • PersonaCraft may generate redundant or conflicting personas despite duplicate-avoidance logic; no mechanism for resolving inter-agent persona conflicts is specified.
  • Scalability of on-demand agent instantiation (latency, token cost) is unaddressed.

Relevance to Harnesses / Meta-Harnesses

This paper is directly relevant to meta-harness design: the orchestrator described is itself a meta-harness — a runtime layer that dynamically assembles and configures sub-agents rather than invoking a fixed pipeline. The four-step orchestration loop (profile → decompose → instantiate → aggregate) maps closely to patterns in harness frameworks that manage agent lifecycle, dependency scheduling, and result synthesis. The session-scoped agent pool with incremental growth is analogous to stateful harness contexts that carry forward intermediate artifacts. For practitioners building meta-harnesses, this paper offers a persona-conditioning abstraction as a lever for user-adaptive orchestration, complementing existing work on static role assignment and fixed coordination graphs.