SCATE framework boosts AI coding agents' test coverage by 32%

Autonomous coding agents have made automated test generation significantly faster, but they suffer from 'lazy generation'—a tendency to prematurely terminate tasks and avoid complex programmatic logic, leading to inadequate code coverage. Currently, mitigating this requires continuous human-in-the-loop supervision, which creates a bottleneck that negates efficiency gains. Researchers from the University of British Columbia (Sijia Gu, Noor Nashid, Ali Mesbah) have introduced SCATE, a framework that replaces human intervention with adaptive, automated supervision of coding agents during test generation.

SCATE formulates supervision as a contextual bandit problem, learning to select the most promising testing actions based on current coverage and class testability metrics. This maximizes coverage gains while minimizing wasted generation effort. Empirically, SCATE integrates seamlessly with different coding agents. When applied to Gemini-CLI, it achieved 32.3% higher line coverage and 30.9% higher branch coverage compared to the agent-only baseline. A comparison with Claude Code confirmed that SCATE dynamically adapts its policy to optimize each agent's unique strengths. The framework also consistently outperformed state-of-the-art non-agentic approaches across all metrics, promising a cost-effective and scalable path to high-quality test generation.