bind_tools, tool calls, and ToolMessage

bind_tools() tells a chat model which tool schemas are available. The model may return requests in AIMessage.tool_calls; it does not execute those Python functions by itself.

The four-step loop

model_with_tools = model.bind_tools([get_order_status])
messages = [{"role": "user", "content": "Where is order A-19?"}]

# 1. Ask the model which action it wants.
ai_message = model_with_tools.invoke(messages)
messages.append(ai_message)

# 2. Validate and execute each request.
for tool_call in ai_message.tool_calls:
    # 3. Tool.invoke(tool_call) returns a matching ToolMessage.
    messages.append(get_order_status.invoke(tool_call))

# 4. Let the model read the result and answer.
answer = model_with_tools.invoke(messages)

A ToolMessage.tool_call_id must match the originating call ID. That pairing matters when the model requests more than one tool.

Failure note

Binding a tool grants the model a vocabulary, not permission. Validate every argument and enforce access at execution time. If you create ToolMessage manually, mismatched or missing call IDs can break the conversation. An agent automates this loop, but the same security boundary remains.

Related

Tools and schemas · create_agent · Retries and backoff