4-bit Quantization Amplifies Tool-Calling Failures in LLM Agents

Quantization to 4-bit weights increases the volume of existing tool-calling failures by up to 2.5x in multi-turn agentic tasks. While standard metrics show negligible change, the error budget masks an amplification of existing hallucinations, such as tool-name errors in specific domains.