Sign InOpen Brain
arXivPaperNeeds Review

Persona-Execution Separation: An Architecture Pattern for Evolving LLM Agents under Execution Audit

Persona-Execution Separation isolates mutable agent behavior from audited, stateful work through a governed bridge. It is an architecture option for multi-user agents under compliance constraints.

arXiv · Aug 27, 2026
Open Source Open MarkdownOpen JSON
Source Summary

**Persona-Execution Separation** places persona and execution in different trust domains joined by a governed contract bridge. Execution remains faceless and audited; summaries can return, while data bodies stay restricted unless a graded DLP exception applies.

Practical Implication

For multi-user agents with frequent prompt or persona changes, keep stateful actions behind typed changes, approvals, DLP, and a stable audit anchor. Treat persona updates as untrusted presentation changes rather than execution-policy changes.

Agent-Ready Context
**Persona-Execution Separation** places persona and execution in different trust domains joined by a governed contract bridge. Execution remains faceless and audited; summaries can return, while data bodies stay restricted unless a graded DLP exception applies.

For multi-user agents with frequent prompt or persona changes, keep stateful actions behind typed changes, approvals, DLP, and a stable audit anchor. Treat persona updates as untrusted presentation changes rather than execution-policy changes.

The evidence is a **one-month pilot** recording five decisions plus checks across **five model configurations**. The pattern is scoped to systems that jointly require persona churn, multi-user deployment, and execution audit.
Connected Context · Feed7 Judgment

This narrows prior harness guidance to agents whose personas change frequently across multiple users: identity and presentation should remain replaceable without altering the trusted execution boundary. It adds a concrete data-flow rule—return summaries by default and expose bodies only through graded DLP exceptions—but its one-month, five-decision pilot across five model configurations supports a scoped architecture pattern, not broad effectiveness claims.

Context Map
agentsecurity#harness-engineering#agent-reliability
Uncertainty
The evidence is a **one-month pilot** recording five decisions plus checks across **five model configurations**. The pattern is scoped to systems that jointly require persona churn, multi-user deployment, and execution audit.