Optimal balance of human–AI interaction modes in ADRS

Determine the optimal balance between synchronous, interactive coding assistants such as Cursor and asynchronous, autonomous algorithm-evolution frameworks such as OpenEvolve as the user interface for AI-Driven Research for Systems, and ascertain when human guidance should be applied versus when autonomous operation is preferable within ADRS pipelines.

Background

The paper proposes AI-Driven Research for Systems (ADRS) as a method to automate solution discovery and evaluation for systems performance problems. Two interaction paradigms are highlighted: synchronous, interactive assistants (e.g., Cursor) that facilitate direct human-in-the-loop guidance, and asynchronous, autonomous frameworks (e.g., OpenEvolve) that iterate solutions without continuous human intervention.

Identifying the most effective interface between researchers and ADRS is crucial for both productivity and solution quality. The authors explicitly note that determining how to balance these modes—deciding when human oversight adds value versus when autonomous search should proceed—is still unresolved, making it a central open question for future ADRS system design and research practice.