Skip to content

Agent Manager

The Agent Manager is the Cat's component that manages the execution of language models chains. A language model chain is a pipeline that takes one or more input variables, it formats them in a prompt, submits the prompt to a language model and, optionally, parses the output.

The Cat's Agent Manager orchestrates two chains:

  1. the tool chain, which, in turn, is a component of the Tool Agent;
  2. the memory chain

When suitable tools for the task at hand are retrieved from the procedural memory, the Agent Manager calls the Tool Agent to execute the tool chain; otherwise the memory chain is executed to answer the user's question with the context retrieved from the episodic and declarative memories.

Schema of the Cheshire Cat memories

Specifically, the default execution pipeline is the following:

  1. the Cat receives the user's message;
  2. the Cat looks for relevant context in each memory collection (i.e. procedural, declarative and episodic) using the user's message as a query;
  3. if meaningful context is retrieved from the procedural memory, the Tool Agent starts, otherwise the memory chain starts;
  4. if executed, the Tool Agent provides an output. If the output answer the user's input, such output is returned to the user, otherwise the memory chain starts;
  5. if executed, the memory chain provides the output using the context retrieved from the declarative and episodic memories.