Right now, tool integration in LLM-powered systems is ad hoc at best. Each agent framework, each plugin system and each model vendor tend to define their own way of handling tool invocation.