SCATE Automates Coding Agent Supervision for Efficient Test Generation.

This paper introduces SCATE, a novel framework designed to automate the supervision of coding agents, specifically for generating software tests. A common challenge with autonomous coding agents is "lazy generation," where they prematurely stop tasks and avoid complex logic, leading to insufficient code coverage. Traditionally, mitigating this requires continuous human intervention, which creates a bottleneck and negates the efficiency benefits of automation. SCATE overcomes this by reframing supervision as a contextual bandit problem. It learns to dynamically select the most promising testing actions based on real-time metrics like current code coverage and class testability. This adaptive approach maximizes coverage gains while minimizing wasted computational effort. Empirical evaluations show that SCATE seamlessly integrates with various coding agents, achieving significantly higher line and branch coverage compared to agent-only baselines and outperforming state-of-the-art non-agentic methods across all metrics.