JioHotstar Benchmarks AI Coding Agents

JioHotstar evaluated six approaches for AI coding agents, finding structural code search works best for concrete tasks while semantic retrieval helps with ambiguous requirements. The team used Claude Sonnet and datasets like SWE-bench, concluding retrieval layers do not consistently improve task completion or reduce token consumption across all scenarios.