Persona-Execution Separation: An Architecture Pattern for Evolving LLM Agents (arXiv 2608.27427)

Forum topic · 小凯 · 2026-08-30

Summary

This arXiv paper (2608.27427) by Yisen Xi, published 2026-08-27 in the machine learning domain, introduces Persona-Execution Separation (PES), an architecture pattern for LLM agents in governed organizations. The core problem: agents need their persona—instructions, tone, self-presentation—to evolve freely, while their execution—stateful, audited work—must remain traceable. A single trust domain cannot satisfy both goals cheaply. PES places persona and execution in separate trust domains connected by a governed contract bridge. The persona component is singly-homed and free to drift; the execution component is faceless and auditable. Only status summaries may cross back into the persona domain, while data bodies stay in the restrictive domain unless covered by a graded data-loss-prevention (DLP) exception; identity remains continuous across the bridge. Crossings are enforced by an approval matrix, DLP policies, and audit mechanisms. The authors derive PES from three goals—free persona drift, execution traceability, and decoupling—and argue, under an assumption of LLM representational indistinguishability, that any single-domain mechanism meeting all three goals must reintroduce typed change objects, external planes, and stable audit anchors, which PES reconstructs at a higher coupling cost.

Paper Overview

Research Area: ML Author: Yisen Xi Published: 2026-08-27 arXiv: 2608.27427

Abstract (English)

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. The paper presents 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 planes, and stable audit anchors: PES rebuilds these at a higher coupling cost.

Key Ideas

Two trust domains: a free-to-evolve persona domain and a restrictive, auditable execution domain.
Governed contract bridge: only status summaries flow back; data bodies stay in the execution domain unless a graded DLP exception applies.
Enforcement mechanisms: approval matrix, DLP, and audit trails control every boundary crossing.
Impossibility-style result: under LLM representational indistinguishability, single-domain designs cannot cheaply meet all three goals (free drift, traceability, decoupling) without re-creating PES-like structure.

---

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