# AWS Bedrock never emits parallel tool calls

medium · inference · probed live 2026-07-26 · affects AWS Bedrock (Chat Completions),
AWS Bedrock (Responses) ·
HTML version: [https://inferencecanary.com/bedrock-parallel-tool-call](https://inferencecanary.com/bedrock-parallel-tool-call)

`parallel_tool_calls: true` is accepted on both Bedrock APIs. Only a single call ever comes
back.

## What was sent

The two-lookup prompt, Chat Completions lane
([https://inferencecanary.com/assets/repro/parallel-calls.chat.json](https://inferencecanary.com/assets/repro/parallel-calls.chat.json)):

```json
{
  "model": "google.gemma-4-31b",
  "messages": [
    {
      "role": "user",
      "content": "Fetch the access codes for BOTH Paris and Tokyo using `get_code` — you may call it twice in parallel. Then report them exactly as: Paris=<code>, Tokyo=<code>."
    }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_code",
        "description": "Returns the access code for a city.",
        "parameters": {
          "type": "object",
          "properties": {
            "city": {
              "type": "string"
            }
          },
          "required": [
            "city"
          ]
        }
      }
    }
  ],
  "parallel_tool_calls": true,
  "temperature": 0,
  "max_tokens": 300
}
```

The Responses-lane variant
([https://inferencecanary.com/assets/repro/parallel-calls.responses.json](https://inferencecanary.com/assets/repro/parallel-calls.responses.json)):

```json
{
  "model": "google.gemma-4-31b",
  "input": [
    {
      "role": "user",
      "content": "Fetch the access codes for BOTH Paris and Tokyo using `get_code` — you may call it twice in parallel. Then report them exactly as: Paris=<code>, Tokyo=<code>."
    }
  ],
  "store": false,
  "tools": [
    {
      "type": "function",
      "name": "get_code",
      "description": "Returns the access code for a city.",
      "parameters": {
        "type": "object",
        "properties": {
          "city": {
            "type": "string"
          }
        },
        "required": [
          "city"
        ]
      }
    }
  ],
  "parallel_tool_calls": true,
  "reasoning": {
    "effort": "none"
  },
  "max_output_tokens": 300
}
```

## What came back

Complete, both lanes:

```json
[
  {
    "id": "call_0",
    "type": "function",
    "function": { "name": "get_code", "arguments": "{\"city\":\"Paris\"}" }
  }
]
```

One call. Paris only — **Tokyo is silently dropped**, with empty content and no error. Every
other provider returns two calls for the identical prompt: Lightning 2, SambaNova 2,
Cerebras 2, Together AI 2, DeepInfra 2, Novita 2, Google 2, Google (OpenAI-compat) 2 — AWS 1,
AWS Responses 1.

## How this breaks expectations

Every other provider proves the model emits parallel calls for this prompt; Bedrock suppresses the
capability at the serving layer while accepting the knob that requests it.

## Why it matters

Fan-out agent loops degrade to serial calling at double the turns and latency — and worse than
serial: here the model reported on Paris and never asked about Tokyo, silently.

## Run it yourself

```
curl "$BASE_URL/chat/completions" -H "Authorization: Bearer $KEY" \
  -H "Content-Type: application/json" -d @parallel-calls.chat.json
```

Two calls back = fixed; one Paris-only call = this finding. The Responses variant POSTs to
`$BASE_URL/responses`.