Parallel Tool Calls in bind_tools

What it does

The langchain-openrouter==0.2.4 update surfaces parallel_tool_calls=True on bind_tools. This tells the underlying LLM it may invoke multiple tool calls in a single response. Models like Claude 3.5, GPT-4o, and Gemini 1.5 support parallel function calling, and the output type switches from a single AIMessage tool call to a list of ToolCall objects inside one AIMessage.

Why it matters

A customer-service agent often needs account balance, recent transactions, and policy details at once. Without parallel tool calls you need sequential chat turns, which are slow, expensive, and error-prone. With the flag, the model fires all three in a single turn, slashing latency and token usage.

Example

$ model = ChatOpenRouter(model="anthropic/claude-3-5-sonnet", openrouter_api_key="your-api-key")
bound = model.bind_tools([WeatherTool, DateTool], parallel_tool_calls=True, strict=True)
resp = bound.invoke([HumanMessage(content="Weather in Tokyo and today's date?")])
print(f"Tool calls: {len(resp.tool_calls)}")
for tc in resp.tool_calls:
    print(f"  Tool: {tc['name']} | Args: {tc['args']}")

Number of tool calls: 2
  Tool: WeatherTool | Args: {'city': 'Tokyo'}
  Tool: DateTool   | Args: {'format': '%Y-%m-%d'}

Common flags

bind_tools

Binds a Pydantic tool schema to a chat model

strict=True

Enforces tool call format validation

parallel_tool_calls

Allows multiple tool calls in one response