Qwen3 agents lose French tool alignment

Hot take: this sounds less like “Qwen got worse at reasoning” and more like the agent lost lexical alignment with the environment it is operating in.

• –If titles, column names, comments, and user questions are French, keeping the core control language in French is a reasonable default for tool selection accuracy.
• –For ReAct/tool agents, the prompt language often matters most where the model maps text to schema, filters, and resource names, not just where it does abstract reasoning.
• –A hybrid setup is usually the safest bet: English for protocol and guardrails, French for examples, schema-matching guidance, and any instructions that depend on the actual data language.
• –Bilingual prompting can help, but it can also add ambiguity if the model starts treating both languages as interchangeable instead of grounding on the dominant one in the data.
• –The most practical test is an A/B eval on the same tool traces: French-only system prompt vs English-only vs hybrid, scored on correct resource/field/filters selection.