Problem

The process_query function in the Python client quickstart (docs/docs/develop/build-client.mdx) has two bugs when the model returns multiple tool calls in a single response:

1. Corrupted conversation history — the for loop appends assistant_message_content to messages and fires a new API call inside the loop, once per tool call. On the second iteration, assistant_message_content has grown and the history is malformed.

2. Crash on non-text first block — after the loop, the code accesses response.content[0].text unconditionally. If the first content block is a tool use block, this raises AttributeError.

Both bugs are silent when only one tool call is returned per turn but surface as soon as the model decides to batch tool calls.

Fix

Replace the for loop with a while response.stop_reason == "tool_use" loop that:

• Appends the full assistant turn (response.content) to history once per turn
• Collects all tool results for that turn before making the next API call
• Falls through to collect the final text response after all tool rounds complete

This matches the agentic loop pattern described in the Anthropic tool use docs.