Bedrock Converse API tool use calls multiple tools in parallel in a single turn

Discovered this while verifying A/B test variant selection with Bedrock Converse API tool use. I defined 3 tool types (get_user_profile, get_similar_users, get_variant_performance) and needed 4 calls total (variant performance for both A and B).

I expected the model to call tools one at a time, inspect results, then call the next. Instead, it requested all 4 in a single response.

Output

Turn 1: get_user_profile({"user_id": "user_001"})
Turn 1: get_similar_users({"user_id": "user_001", "limit": 10})
Turn 1: get_variant_performance({"experiment_id": "cta_test_2024", "variant_id": "A"})
Turn 1: get_variant_performance({"experiment_id": "cta_test_2024", "variant_id": "B"})

All requested in Turn 1. After returning results, the model gave its final decision in Turn 2. Every verification completed in exactly 2 turns.

The implementation detail: output_message["content"] can contain multiple toolUse blocks, so you need to loop through all of them and return all toolResults together.

Python

if response["stopReason"] == "tool_use":
    tool_results = []
    for block in output_message["content"]:
        if "toolUse" in block:
            tool = block["toolUse"]
            result = execute_tool(tool["name"], tool["input"])
            tool_results.append({"toolResult": {
                "toolUseId": tool["toolUseId"],
                "content": [{"json": json.loads(result)}],
            }})
    messages.append({"role": "user", "content": tool_results})

The tool use documentation sample code focuses on single-tool examples, though the for loop structure does handle multiple toolUse blocks. When tools have no dependencies, the model is likely to choose parallel calls.