ToolCallingAgents do not follow the ReAct (Yao et al., 2022) Paradigm

[michaeltheologitis]

Hello everyone!

The documentation implies that smolagent agents work with the ReAct (Yao et al., 2022) paradigm. This is true only for the CodeAgents and not on the ToolCallingAgents — you can validate this from the specific prompts at:
https://github.com/huggingface/smolagents/tree/main/src/smolagents/prompts

The ToolCallingAgents work by simple Actions/Observation (without Reasoning or any CoT). This was unclear in the docs and I am wondering: 1. Is this intentional?, and 2. Why is this happening when we can have better results by simply adding reasoning to the prompt requirements?

[KeepALifeUS]

Great observation! This is an important distinction.

My take on why this might be intentional:

1. Token efficiency trade-off: ReAct adds reasoning tokens before each action. For simple tool calls, this can be overhead without much benefit. The model already "reasons" internally before generating the tool call.

2. Model capability: Modern tool-calling APIs (OpenAI, Anthropic) are specifically trained for structured tool use. They may not need explicit CoT prompting the way earlier models did when ReAct was developed.

However, I agree reasoning helps in complex scenarios:

When I build multi-agent systems, I find explicit reasoning valuable for:

• Multi-step planning — agent needs to reason about dependencies
• Debugging — seeing the reasoning trace helps understand failures
• Complex tool selection — when multiple tools could apply

Alternative approach I've found effective:

Instead of in-prompt reasoning, use state-based reasoning:

state = {
    "reasoning": "Need to fetch data before analysis",
    "next_action": "fetch_data",
    "rationale": {}
}

The agent writes its reasoning to shared state, which:

• Makes reasoning observable and debuggable
• Doesn't bloat the conversation context
• Allows other agents to understand the decision

This is similar to what Anthropic calls "thinking" tokens but externalized.

Working example: https://github.com/KeepALifeUS/autonomous-agents

So to answer your question: I suspect it's a pragmatic trade-off, but for complex workflows, adding structured reasoning (either in prompt or in state) definitely improves results.