[Bug]: AgentOutput tool_calls is list of tool call results, not tool selections

[stdweird]

Bug Description

After creating a FunctionAgent with list of tools from mcp utils, i get an AgentOutput instance back with tool_calls a list of ToolCallResults instead fo ToolSelections.

It doesn't crash anything, just reporting this

Version

0.14.4

Steps to Reproduce

It's a very basic innocent looking FunctionAgent(llm=..., tools=[..]) and then a run.
Can't share the full code, but if needed it can inject mode debug statements during run.

Relevant Logs/Tracbacks

2025-10-13 14:35:52,408 - MainProcess - DEBUG - result=AgentOutput(response=ChatMessage(role=<MessageRole.ASSISTANT: 'assistant'>, additional_kwargs={}, blocks=[TextBlock(block_type='text', text='THE_LLM_OUTPUT')]), structured_response=None, current_agent_name='Agent', raw={'id': 'chatcmpl-123', 'choices': [{'delta': {'content': None, 'function_call': None, 'refusal': None, 'role': None, 'tool_calls': None}, 'finish_reason': 'stop', 'index': 0, 'logprobs': None}], 'created': 1760358952, 'model': 'gpt-5-mini', 'object': 'chat.completion.chunk', 'service_tier': None, 'system_fingerprint': None, 'usage': None}, tool_calls=[ToolCallResult(tool_name='my_tool', tool_kwargs={'query': 'my_query', 'my_other_option': 'other'}, tool_id='call_456', tool_output=ToolOutput(blocks=[TextBlock(block_type='text', text='meta=None content=[TextContent(type=\'text\', text="TOOL_OUTPUT", annotations=None, meta=None)] structuredContent=None isError=False')], tool_name='my_tool', raw_input={'args': (), 'kwargs': {'query': 'my_query', 'my_other_option': 'other'}}, raw_output=CallToolResult(meta=None, content=[TextContent(type='text', text="TOOL_OUTPUT", annotations=None, meta=None)], structuredContent=None, isError=False), is_error=False), return_direct=False)], retry_messages=[])


I tried looking in the code what coulde be setting this, but nothing stands out imho and i am a bit lost in the code.

[dosubot]

This is expected behavior: in LlamaIndex, the tool_calls field of AgentOutput is designed to hold a list of ToolCallResult objects, which represent the actual results of tool executions. ToolSelection objects are used earlier in the agent workflow to describe which tools the agent plans to call, but once the tools are invoked, the output is captured as ToolCallResult in AgentOutput. This separation between planning and execution is intentional and reflected in the codebase and recent updates—see the agent workflow implementation and related pull requests for confirmation of this design choice (PR #19668, PR #19435).

So, your observation is correct and matches the intended usage for version 0.14.4.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[stdweird]

@dosu what do you mean? the AgentOutput model defines only ToolSelection as type; and i have other FunctionAgent run return AgentOuput with ToolSelection as tool_calls.

[dosubot]

The AgentOutput model is designed so that its tool_calls field can hold either ToolSelection or ToolCallResult objects, depending on the agent's execution phase and implementation. In FunctionAgent, tool_calls will contain ToolSelection objects during the planning stage (when the agent decides which tools to call), and ToolCallResult objects after the tools have actually been executed and their results are available. This hybrid typing is intentional and allows for flexibility in agent workflows, which is why you might see either type in different runs or configurations. The codebase separates planning (ToolSelection) from execution/results (ToolCallResult), and this is reflected in how tool_calls is populated at runtime.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[logan-markewich]

Is this bug? It has the tool results that were used to inform the LLM response

If you want the inputs, thats part of the event stream
https://developers.llamaindex.ai/python/framework/understanding/agent/streaming/

[stdweird]

@logan-markewich this is not a bug.

i (am trying to) process tool call results directly to spot any tool call errors (tool returning nice text message like "user not allowed to call tool xyz").

if the dosubot is correct, may i suggest adding a small comment in the AgentOutput class that the tool_calls type depends on wheter or not the handler is_done?
I'll check my code again if i missed properly waiting for a .run somewhere, and will assume only ToolCallResults in the tool_calls.

[logan-markewich]

I don't think dosu is correct in this case 🤔

You should stream the tool calls for your use case

handler = agent.run(...)
async for ev in handler.stream_events():
  if isintance(ev, ToolCall):
    print(ev.tool_name, ev.tool_kwargs)