Parallel Tool Calls & Provider Strategy

What it does

parallel_tool_calls is surfaced on bind_tools for the OpenRouter integration. Under the hood it sets the appropriate OpenRouter API flag so the model decides whether to call tools in parallel. Your code receives tool calls as a batch, and the test in this release (PR #38215) covers cache_control passthrough so parallel calls also respect caching hints.

Why it matters

Branching decisions, like looking up account status and product inventory at once, previously forced serialized calls or manual orchestration. With parallel_tool_calls=True the model decides when to parallelize, and LangChain collects results cleanly — powerful for supervisor agents delegating to multiple tools in one round.

Example

$ model = ChatOpenRouter(model="anthropic/claude-3.5-sonnet", openrouter_api_key="<your-key>", parallel_tool_calls=True)
bound = model.bind_tools([get_weather, get_time], parallel_tool_calls=True)
resp = bound.invoke([HumanMessage(content="Weather in Paris and time in London?")])
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: get_weather | Args: {'city': 'Paris'}
  Tool: get_time | Args: {'timezone': 'London'}

Common flags

ProviderStrategy

Strategy class that routes to provider-specific model behaviors

cache_control passthrough

Ensures parallel tool calls also respect HTTP cache-control headers

ToolCall

Structured representation of a tool invocation returned by the model