feat: Thinking LLMs in agents

[jakubduda-dsai]

Feature description

Introduce support for "Thinking LLMs" inside agents. The agent should be able to run an internal reasoning step (hidden chain of thought) before producing the result. This could be implemented by wrapping the model calls so that the agent can:

• Generate reasoning tokens separately from the final output.
• Optionally persist or log these reasoning traces for debugging, benchmarking, or evaluation.
• Allow to configure whether the "thinking" step remains hidden or visible

Motivation

By integrating Thinking LLMs:

• We can improve reasoning quality and consistency of agent behavior.
• Developers gain better insight into why the agent took certain steps (debugging, error analysis).
• It aligns with modern approaches to tool-using and planning agents, where separating internal thought from final answers improves reliability.

Additional context

No response