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