Investigate decomposition & Chaining

[michael-desmond]

Is your feature request related to a problem? Please describe.

Rather than relying on a single agent to perform a sequence of actions (such as determining if a tool is necessary, calling the tool, determining if enough information is available (maybe call another tool), producing a response etc. ) we may be able to produce more reliable (and configurable) agents if we could deconstruct this overall behavior into an set of independent operations and chain them together.

This is particularly interesting when considering smaller and relatively less capable LLMs.

Describe the solution you'd like

We already have some of the core pieces to support decomposition such as the ChatLLM with structured output, the tools framework etc. However how to chain one operator into another, possibly with conditional routing, is still unclear. The most obvious solution would be to perhaps introduce a graph construct.