Orchestration of isolated LLM roles as an alternative to manually switching between chatbots

The author of the article on Root.cz (the second installment in a series on working with language models) describes the shift from manually alternating between chatbots to automated orchestration. In the previous installment, he showed how he manually assigns a task to one model, transfers the result to a second for critical review, and returns the feedback to the first – the process works, but requires a person between individual steps.

The proposed concept replaces manual switching with a software agent that receives a single task (for example, “propose a business model and critically review it”) and internally breaks it down into separate roles – author, economic review, technical review, legal review, contradiction checking, revision, final document. From the user’s perspective, this is one request; internally, multiple roles work in coordination.

The author’s key finding is that context isolation between roles does not depend on whether the roles are handled by different model providers. The same effect can be achieved through several independent calls to the same model, provided each role receives a new, fresh request without the history of the previous conversation – according to the author, context is what matters, not the provider’s name. In such an architecture, a smaller local model can act as a “dispatcher” that decides where to route the task (request classification, tool selection, context preparation, document shortening, etc.), while the external model is called only where its capability is actually needed.

In practice, however, the author still prefers using multiple providers simultaneously, because he believes different models have different strengths and blind spots, and combining them reduces the risk of the entire process remaining confined to a single way of thinking. He proposes an architecture consisting of a “local orchestrator plus a set of available capabilities”, which includes non-language tools alongside various LLMs – for calculations, he mentions MATLAB, Mathematica, Mathcad or Python, for example, instead of simulating calculations with a language model. The full text of the source article is not available; details can be found in the source article.