Hey guys, I am looking into some leaks of reasoning with bedrock agents from Anthropic where the agent speaks another language, switches to English, and I can hear it reason out loud (saying stuff like, I need to say this call that tool etc..) and then switch to the original lang of the conversation and continuing. Is this a known issue? Did anyone experience this before? anything shared on this before ?

Try updating your system prompt to explicitly prevent the agent from reasoning out loud. Add a line like this to your prompt:

Do not verbalize your internal reasoning process. Only speak responses intended for the user in their language.

This helps control multilingual behavior and keeps internal thoughts silent, as recommended in the [multilingual configuration guide](https://docs.vapi.ai/customization/multilingual). Source: - [Multilingual support documentation](https://docs.vapi.ai/customization/multilingual)

Yes, I’ve seen this before with Anthropic agents on Bedrock, especially in multilingual voice setups. It’s usually caused by internal reasoning/tool-planning leaking into the user-facing stream, so the model accidentally says things like “I need to call this tool…” before continuing normally. Could be from the orchestration layer, streaming handler, or TTS pipeline exposing hidden chain-of-thought tokens. I can help debug it if you want. @Tancho

I just wanna handle it. And it's.not sharing thoughts

Let's connect @Tancho