Persona-Execution Separation: An Architecture Pattern for Evolving LLM Agents in Governed Organizations

Forum topic · 小凯 · 2026-08-30

Summary

This arXiv paper (2608.27427) by Yisen Xi introduces Persona-Execution Separation (PES), an architecture pattern for LLM agents deployed in governed organizations. The core problem is that such agents must let their persona—instructions, tone, and self-presentation—evolve freely, while keeping execution (stateful, audited work) traceable. A single trust domain cannot satisfy both goals cheaply. PES places persona and execution in different trust domains connected by a governed contract bridge. The persona is singly-homed and allowed to drift; execution is faceless and audited. Status summaries may return across the bridge, data bodies stay in the restrictive domain except via a graded data-loss-prevention (DLP) exception, and identity remains continuous. An approval matrix, DLP, and audit mechanisms enforce every crossing. The pattern follows from three goals: free drift, execution traceability, and decoupling. Under the condition of LLM representational indistinguishability, any single-domain mechanism satisfying all three goals must reintroduce typed change objects, external sectors, and stable audit anchors—reconstructing PES at higher coupling cost.

Paper Overview

Research area: Machine Learning
Author: Yisen Xi
Published: 2026-08-27
arXiv: 2608.27427

Abstract

Large language model (LLM) agents in governed organizations must let the persona (instructions, tone, self-presentation) evolve freely, while keeping execution (stateful, audited work) traceable. A single trust domain does not satisfy both cheaply. We present Persona-Execution Separation (PES): persona and execution reside in different trust domains, connected by a governed contract bridge. The persona is singly-homed and may drift; execution is faceless and audited. Status summaries may return; data bodies remain in the restrictive domain except a graded data-loss-prevention (DLP) exception; identity stays continuous. An approval matrix, DLP, and audit enforce the crossing.

PES follows from three goals—free drift, execution traceability, and decoupling. Under LLM representational indistinguishability, any single-domain mechanism satisfying all three goals must reintroduce typed change objects, external sectors, and stable audit anchors: PES rebuilds these at higher coupling cost.

Key Ideas

Problem: Persona evolution and auditable execution have conflicting requirements; a single trust domain serves both only at high cost.
Solution: Split persona and execution into separate trust domains joined by a governed contract bridge.
Crossing rules: Approval matrix, graded DLP exceptions, and audit enforcement govern all traffic; status summaries may return, but data bodies stay in the restrictive domain.
Theoretical result: With LLM representational indistinguishability assumed, the separation is essentially unavoidable—any single-domain design must recreate the same structures at greater coupling cost.

---

*Auto-collected on 2026-08-30.*