All publications

Industrial LLM-based code optimization under regulation: a mixture-of-agents approach

2025·arXiv
compliance boundary — approved models · full audit logagent Aagent Bagent Cmixturereview · combinevalidatelog every stepaudited changeonly validated, logged changes cross the boundary

Figure 1: Optimization inside the boundary. A mixture of approved agents generates and reviews candidates entirely within the compliance perimeter (dashed); only validated, logged changes cross it.

Regulated estates cannot ship code to arbitrary model endpoints, and they cannot accept changes without an audit trail. This paper describes a mixture-of-agents architecture for code optimization that operates inside a compliance boundary: approved models only, every candidate logged, every change validated before it crosses the line.

The mixture matters because no approved model is strongest at everything; combining agents recovers much of the capability lost by excluding frontier endpoints, while the boundary preserves the guarantees regulators require.

Key results

  • Mixture-of-agents recovers capability using only approved models
  • Full audit trail: every candidate and decision is logged inside the boundary
  • Designed for on-premise and air-gapped deployment