Use native tool calling instead of injecting tool description in agent context

[c0dezli]

Currently, we inject tool descriptions into the agent context for all LLM interactions. While this approach works, it doesn't take advantage of the native tool calling capabilities that modern LLMs like Claude and GPT provide.

I propose for LLMs with native tool calling support (like Claude and GPT), we should use litellm's tool interface to directly register tools and leverage their native function calling formats. This would allow us to remove tool descriptions from the agent context for these models.

For LLMs without native tool support, we can maintain the current approach of injecting tool descriptions in the agent context.